thys/Journal/Paper.thy
changeset 218 16af5b8bd285
child 265 d36be1e356c0
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/thys/Journal/Paper.thy	Sun Feb 26 23:46:22 2017 +0000
@@ -0,0 +1,1260 @@
+(*<*)
+theory Paper
+imports 
+   "../Lexer"
+   "../Simplifying" 
+   "../Sulzmann" 
+   "~~/src/HOL/Library/LaTeXsugar"
+begin
+
+declare [[show_question_marks = false]]
+
+abbreviation 
+ "der_syn r c \<equiv> der c r"
+
+abbreviation 
+ "ders_syn r s \<equiv> ders s r"
+
+notation (latex output)
+  If  ("(\<^raw:\textrm{>if\<^raw:}> (_)/ \<^raw:\textrm{>then\<^raw:}> (_)/ \<^raw:\textrm{>else\<^raw:}> (_))" 10) and
+  Cons ("_\<^raw:\mbox{$\,$}>::\<^raw:\mbox{$\,$}>_" [75,73] 73) and  
+
+  ZERO ("\<^bold>0" 78) and 
+  ONE ("\<^bold>1" 78) and 
+  CHAR ("_" [1000] 80) and
+  ALT ("_ + _" [77,77] 78) and
+  SEQ ("_ \<cdot> _" [77,77] 78) and
+  STAR ("_\<^sup>\<star>" [1000] 78) and
+  
+  val.Void ("'(')" 1000) and
+  val.Char ("Char _" [1000] 78) and
+  val.Left ("Left _" [79] 78) and
+  val.Right ("Right _" [1000] 78) and
+  val.Seq ("Seq _ _" [79,79] 78) and
+  val.Stars ("Stars _" [79] 78) and
+
+  L ("L'(_')" [10] 78) and
+  der_syn ("_\\_" [79, 1000] 76) and  
+  ders_syn ("_\\_" [79, 1000] 76) and
+  flat ("|_|" [75] 74) and
+  Sequ ("_ @ _" [78,77] 63) and
+  injval ("inj _ _ _" [79,77,79] 76) and 
+  mkeps ("mkeps _" [79] 76) and 
+  length ("len _" [73] 73) and
+ 
+  Prf ("_ : _" [75,75] 75) and  
+  Posix ("'(_, _') \<rightarrow> _" [63,75,75] 75) and
+ 
+  lexer ("lexer _ _" [78,78] 77) and 
+  F_RIGHT ("F\<^bsub>Right\<^esub> _") and
+  F_LEFT ("F\<^bsub>Left\<^esub> _") and  
+  F_ALT ("F\<^bsub>Alt\<^esub> _ _") and
+  F_SEQ1 ("F\<^bsub>Seq1\<^esub> _ _") and
+  F_SEQ2 ("F\<^bsub>Seq2\<^esub> _ _") and
+  F_SEQ ("F\<^bsub>Seq\<^esub> _ _") and
+  simp_SEQ ("simp\<^bsub>Seq\<^esub> _ _" [1000, 1000] 1) and
+  simp_ALT ("simp\<^bsub>Alt\<^esub> _ _" [1000, 1000] 1) and
+  slexer ("lexer\<^sup>+" 1000) and
+
+  ValOrd ("_ >\<^bsub>_\<^esub> _" [77,77,77] 77) and
+  ValOrdEq ("_ \<ge>\<^bsub>_\<^esub> _" [77,77,77] 77)
+
+definition 
+  "match r s \<equiv> nullable (ders s r)"
+
+(*
+comments not implemented
+
+p9. The condtion "not exists s3 s4..." appears often enough (in particular in
+the proof of Lemma 3) to warrant a definition.
+
+*)
+
+(*>*)
+
+section {* Introduction *}
+
+
+text {*
+
+Brzozowski \cite{Brzozowski1964} introduced the notion of the {\em
+derivative} @{term "der c r"} of a regular expression @{text r} w.r.t.\ a
+character~@{text c}, and showed that it gave a simple solution to the
+problem of matching a string @{term s} with a regular expression @{term r}:
+if the derivative of @{term r} w.r.t.\ (in succession) all the characters of
+the string matches the empty string, then @{term r} matches @{term s} (and
+{\em vice versa}). The derivative has the property (which may almost be
+regarded as its specification) that, for every string @{term s} and regular
+expression @{term r} and character @{term c}, one has @{term "cs \<in> L(r)"} if
+and only if \mbox{@{term "s \<in> L(der c r)"}}. The beauty of Brzozowski's
+derivatives is that they are neatly expressible in any functional language,
+and easily definable and reasoned about in theorem provers---the definitions
+just consist of inductive datatypes and simple recursive functions. A
+mechanised correctness proof of Brzozowski's matcher in for example HOL4
+has been mentioned by Owens and Slind~\cite{Owens2008}. Another one in Isabelle/HOL is part
+of the work by Krauss and Nipkow \cite{Krauss2011}. And another one in Coq is given
+by Coquand and Siles \cite{Coquand2012}.
+
+If a regular expression matches a string, then in general there is more than
+one way of how the string is matched. There are two commonly used
+disambiguation strategies to generate a unique answer: one is called GREEDY
+matching \cite{Frisch2004} and the other is POSIX
+matching~\cite{Kuklewicz,Sulzmann2014,Vansummeren2006}. For example consider
+the string @{term xy} and the regular expression \mbox{@{term "STAR (ALT
+(ALT x y) xy)"}}. Either the string can be matched in two `iterations' by
+the single letter-regular expressions @{term x} and @{term y}, or directly
+in one iteration by @{term xy}. The first case corresponds to GREEDY
+matching, which first matches with the left-most symbol and only matches the
+next symbol in case of a mismatch (this is greedy in the sense of preferring
+instant gratification to delayed repletion). The second case is POSIX
+matching, which prefers the longest match.
+
+In the context of lexing, where an input string needs to be split up into a
+sequence of tokens, POSIX is the more natural disambiguation strategy for
+what programmers consider basic syntactic building blocks in their programs.
+These building blocks are often specified by some regular expressions, say
+@{text "r\<^bsub>key\<^esub>"} and @{text "r\<^bsub>id\<^esub>"} for recognising keywords and
+identifiers, respectively. There are two underlying (informal) rules behind
+tokenising a string in a POSIX fashion according to a collection of regular
+expressions:
+
+\begin{itemize} 
+\item[$\bullet$] \emph{The Longest Match Rule} (or \emph{``maximal munch rule''}):
+The longest initial substring matched by any regular expression is taken as
+next token.\smallskip
+
+\item[$\bullet$] \emph{Priority Rule:}
+For a particular longest initial substring, the first regular expression
+that can match determines the token.
+\end{itemize}
+
+\noindent Consider for example a regular expression @{text "r\<^bsub>key\<^esub>"} for recognising keywords
+such as @{text "if"}, @{text "then"} and so on; and @{text "r\<^bsub>id\<^esub>"}
+recognising identifiers (say, a single character followed by
+characters or numbers).  Then we can form the regular expression
+@{text "(r\<^bsub>key\<^esub> + r\<^bsub>id\<^esub>)\<^sup>\<star>"} and use POSIX matching to tokenise strings,
+say @{text "iffoo"} and @{text "if"}.  For @{text "iffoo"} we obtain
+by the Longest Match Rule a single identifier token, not a keyword
+followed by an identifier. For @{text "if"} we obtain by the Priority
+Rule a keyword token, not an identifier token---even if @{text "r\<^bsub>id\<^esub>"}
+matches also.
+
+One limitation of Brzozowski's matcher is that it only generates a
+YES/NO answer for whether a string is being matched by a regular
+expression.  Sulzmann and Lu~\cite{Sulzmann2014} extended this matcher
+to allow generation not just of a YES/NO answer but of an actual
+matching, called a [lexical] {\em value}. They give a simple algorithm
+to calculate a value that appears to be the value associated with
+POSIX matching.  The challenge then is to specify that value, in an
+algorithm-independent fashion, and to show that Sulzmann and Lu's
+derivative-based algorithm does indeed calculate a value that is
+correct according to the specification.
+
+The answer given by Sulzmann and Lu \cite{Sulzmann2014} is to define a
+relation (called an ``order relation'') on the set of values of @{term
+r}, and to show that (once a string to be matched is chosen) there is
+a maximum element and that it is computed by their derivative-based
+algorithm. This proof idea is inspired by work of Frisch and Cardelli
+\cite{Frisch2004} on a GREEDY regular expression matching
+algorithm. However, we were not able to establish transitivity and
+totality for the ``order relation'' by Sulzmann and Lu. In Section
+\ref{argu} we identify some inherent problems with their approach (of
+which some of the proofs are not published in \cite{Sulzmann2014});
+perhaps more importantly, we give a simple inductive (and
+algorithm-independent) definition of what we call being a {\em POSIX
+value} for a regular expression @{term r} and a string @{term s}; we
+show that the algorithm computes such a value and that such a value is
+unique. Our proofs are both done by hand and checked in Isabelle/HOL.  The
+experience of doing our proofs has been that this mechanical checking
+was absolutely essential: this subject area has hidden snares. This
+was also noted by Kuklewicz \cite{Kuklewicz} who found that nearly all
+POSIX matching implementations are ``buggy'' \cite[Page
+203]{Sulzmann2014} and by Grathwohl et al \cite[Page 36]{CrashCourse2014}
+who wrote:
+
+\begin{quote}
+\it{}``The POSIX strategy is more complicated than the greedy because of 
+the dependence on information about the length of matched strings in the 
+various subexpressions.''
+\end{quote}
+
+%\footnote{The relation @{text "\<ge>\<^bsub>r\<^esub>"} defined by Sulzmann and Lu \cite{Sulzmann2014} 
+%is a relation on the
+%values for the regular expression @{term r}; but it only holds between
+%@{term "v\<^sub>1"} and @{term "v\<^sub>2"} in cases where @{term "v\<^sub>1"} and @{term "v\<^sub>2"} have
+%the same flattening (underlying string). So a counterexample to totality is
+%given by taking two values @{term "v\<^sub>1"} and @{term "v\<^sub>2"} for @{term r} that
+%have different flattenings (see Section~\ref{posixsec}). A different
+%relation @{text "\<ge>\<^bsub>r,s\<^esub>"} on the set of values for @{term r}
+%with flattening @{term s} is definable by the same approach, and is indeed
+%total; but that is not what Proposition 1 of \cite{Sulzmann2014} does.}
+
+
+\noindent {\bf Contributions:} We have implemented in Isabelle/HOL the
+derivative-based regular expression matching algorithm of
+Sulzmann and Lu \cite{Sulzmann2014}. We have proved the correctness of this
+algorithm according to our specification of what a POSIX value is (inspired
+by work of Vansummeren \cite{Vansummeren2006}). Sulzmann
+and Lu sketch in \cite{Sulzmann2014} an informal correctness proof: but to
+us it contains unfillable gaps.\footnote{An extended version of
+\cite{Sulzmann2014} is available at the website of its first author; this
+extended version already includes remarks in the appendix that their
+informal proof contains gaps, and possible fixes are not fully worked out.}
+Our specification of a POSIX value consists of a simple inductive definition
+that given a string and a regular expression uniquely determines this value.
+Derivatives as calculated by Brzozowski's method are usually more complex
+regular expressions than the initial one; various optimisations are
+possible. We prove the correctness when simplifications of @{term "ALT ZERO
+r"}, @{term "ALT r ZERO"}, @{term "SEQ ONE r"} and @{term "SEQ r ONE"} to
+@{term r} are applied.
+
+*}
+
+section {* Preliminaries *}
+
+text {* \noindent Strings in Isabelle/HOL are lists of characters with the
+empty string being represented by the empty list, written @{term "[]"}, and
+list-cons being written as @{term "DUMMY # DUMMY"}. Often we use the usual
+bracket notation for lists also for strings; for example a string consisting
+of just a single character @{term c} is written @{term "[c]"}. By using the
+type @{type char} for characters we have a supply of finitely many
+characters roughly corresponding to the ASCII character set. Regular
+expressions are defined as usual as the elements of the following inductive
+datatype:
+
+  \begin{center}
+  @{text "r :="}
+  @{const "ZERO"} $\mid$
+  @{const "ONE"} $\mid$
+  @{term "CHAR c"} $\mid$
+  @{term "ALT r\<^sub>1 r\<^sub>2"} $\mid$
+  @{term "SEQ r\<^sub>1 r\<^sub>2"} $\mid$
+  @{term "STAR r"} 
+  \end{center}
+
+  \noindent where @{const ZERO} stands for the regular expression that does
+  not match any string, @{const ONE} for the regular expression that matches
+  only the empty string and @{term c} for matching a character literal. The
+  language of a regular expression is also defined as usual by the
+  recursive function @{term L} with the six clauses:
+
+  \begin{center}
+  \begin{tabular}{l@ {\hspace{3mm}}rcl}
+  (1) & @{thm (lhs) L.simps(1)} & $\dn$ & @{thm (rhs) L.simps(1)}\\
+  (2) & @{thm (lhs) L.simps(2)} & $\dn$ & @{thm (rhs) L.simps(2)}\\
+  (3) & @{thm (lhs) L.simps(3)} & $\dn$ & @{thm (rhs) L.simps(3)}\\
+  \end{tabular}
+  \hspace{14mm}
+  \begin{tabular}{l@ {\hspace{3mm}}rcl}
+  (4) & @{thm (lhs) L.simps(4)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) L.simps(4)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  (5) & @{thm (lhs) L.simps(5)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) L.simps(5)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  (6) & @{thm (lhs) L.simps(6)} & $\dn$ & @{thm (rhs) L.simps(6)}\\
+  \end{tabular}
+  \end{center}
+  
+  \noindent In clause (4) we use the operation @{term "DUMMY ;;
+  DUMMY"} for the concatenation of two languages (it is also list-append for
+  strings). We use the star-notation for regular expressions and for
+  languages (in the last clause above). The star for languages is defined
+  inductively by two clauses: @{text "(i)"} the empty string being in
+  the star of a language and @{text "(ii)"} if @{term "s\<^sub>1"} is in a
+  language and @{term "s\<^sub>2"} in the star of this language, then also @{term
+  "s\<^sub>1 @ s\<^sub>2"} is in the star of this language. It will also be convenient
+  to use the following notion of a \emph{semantic derivative} (or \emph{left
+  quotient}) of a language defined as
+  %
+  \begin{center}
+  @{thm Der_def}\;.
+  \end{center}
+ 
+  \noindent
+  For semantic derivatives we have the following equations (for example
+  mechanically proved in \cite{Krauss2011}):
+  %
+  \begin{equation}\label{SemDer}
+  \begin{array}{lcl}
+  @{thm (lhs) Der_null}  & \dn & @{thm (rhs) Der_null}\\
+  @{thm (lhs) Der_empty}  & \dn & @{thm (rhs) Der_empty}\\
+  @{thm (lhs) Der_char}  & \dn & @{thm (rhs) Der_char}\\
+  @{thm (lhs) Der_union}  & \dn & @{thm (rhs) Der_union}\\
+  @{thm (lhs) Der_Sequ}  & \dn & @{thm (rhs) Der_Sequ}\\
+  @{thm (lhs) Der_star}  & \dn & @{thm (rhs) Der_star}
+  \end{array}
+  \end{equation}
+
+
+  \noindent \emph{\Brz's derivatives} of regular expressions
+  \cite{Brzozowski1964} can be easily defined by two recursive functions:
+  the first is from regular expressions to booleans (implementing a test
+  when a regular expression can match the empty string), and the second
+  takes a regular expression and a character to a (derivative) regular
+  expression:
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) nullable.simps(1)} & $\dn$ & @{thm (rhs) nullable.simps(1)}\\
+  @{thm (lhs) nullable.simps(2)} & $\dn$ & @{thm (rhs) nullable.simps(2)}\\
+  @{thm (lhs) nullable.simps(3)} & $\dn$ & @{thm (rhs) nullable.simps(3)}\\
+  @{thm (lhs) nullable.simps(4)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) nullable.simps(4)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) nullable.simps(5)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) nullable.simps(5)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) nullable.simps(6)} & $\dn$ & @{thm (rhs) nullable.simps(6)}\medskip\\
+  
+  %\end{tabular}
+  %\end{center}
+
+  %\begin{center}
+  %\begin{tabular}{lcl}
+  
+  @{thm (lhs) der.simps(1)} & $\dn$ & @{thm (rhs) der.simps(1)}\\
+  @{thm (lhs) der.simps(2)} & $\dn$ & @{thm (rhs) der.simps(2)}\\
+  @{thm (lhs) der.simps(3)} & $\dn$ & @{thm (rhs) der.simps(3)}\\
+  @{thm (lhs) der.simps(4)[of c "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) der.simps(4)[of c "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) der.simps(6)} & $\dn$ & @{thm (rhs) der.simps(6)}
+  \end{tabular}
+  \end{center}
+ 
+  \noindent
+  We may extend this definition to give derivatives w.r.t.~strings:
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) ders.simps(1)} & $\dn$ & @{thm (rhs) ders.simps(1)}\\
+  @{thm (lhs) ders.simps(2)} & $\dn$ & @{thm (rhs) ders.simps(2)}\\
+  \end{tabular}
+  \end{center}
+
+  \noindent Given the equations in \eqref{SemDer}, it is a relatively easy
+  exercise in mechanical reasoning to establish that
+
+  \begin{proposition}\label{derprop}\mbox{}\\ 
+  \begin{tabular}{ll}
+  @{text "(1)"} & @{thm (lhs) nullable_correctness} if and only if
+  @{thm (rhs) nullable_correctness}, and \\ 
+  @{text "(2)"} & @{thm[mode=IfThen] der_correctness}.
+  \end{tabular}
+  \end{proposition}
+
+  \noindent With this in place it is also very routine to prove that the
+  regular expression matcher defined as
+  %
+  \begin{center}
+  @{thm match_def}
+  \end{center}
+
+  \noindent gives a positive answer if and only if @{term "s \<in> L r"}.
+  Consequently, this regular expression matching algorithm satisfies the
+  usual specification for regular expression matching. While the matcher
+  above calculates a provably correct YES/NO answer for whether a regular
+  expression matches a string or not, the novel idea of Sulzmann and Lu
+  \cite{Sulzmann2014} is to append another phase to this algorithm in order
+  to calculate a [lexical] value. We will explain the details next.
+
+*}
+
+section {* POSIX Regular Expression Matching\label{posixsec} *}
+
+text {* 
+
+  The clever idea by Sulzmann and Lu \cite{Sulzmann2014} is to define
+  values for encoding \emph{how} a regular expression matches a string
+  and then define a function on values that mirrors (but inverts) the
+  construction of the derivative on regular expressions. \emph{Values}
+  are defined as the inductive datatype
+
+  \begin{center}
+  @{text "v :="}
+  @{const "Void"} $\mid$
+  @{term "val.Char c"} $\mid$
+  @{term "Left v"} $\mid$
+  @{term "Right v"} $\mid$
+  @{term "Seq v\<^sub>1 v\<^sub>2"} $\mid$ 
+  @{term "Stars vs"} 
+  \end{center}  
+
+  \noindent where we use @{term vs} to stand for a list of
+  values. (This is similar to the approach taken by Frisch and
+  Cardelli for GREEDY matching \cite{Frisch2004}, and Sulzmann and Lu
+  for POSIX matching \cite{Sulzmann2014}). The string underlying a
+  value can be calculated by the @{const flat} function, written
+  @{term "flat DUMMY"} and defined as:
+
+  \begin{center}
+  \begin{tabular}[t]{lcl}
+  @{thm (lhs) flat.simps(1)} & $\dn$ & @{thm (rhs) flat.simps(1)}\\
+  @{thm (lhs) flat.simps(2)} & $\dn$ & @{thm (rhs) flat.simps(2)}\\
+  @{thm (lhs) flat.simps(3)} & $\dn$ & @{thm (rhs) flat.simps(3)}\\
+  @{thm (lhs) flat.simps(4)} & $\dn$ & @{thm (rhs) flat.simps(4)}
+  \end{tabular}\hspace{14mm}
+  \begin{tabular}[t]{lcl}
+  @{thm (lhs) flat.simps(5)[of "v\<^sub>1" "v\<^sub>2"]} & $\dn$ & @{thm (rhs) flat.simps(5)[of "v\<^sub>1" "v\<^sub>2"]}\\
+  @{thm (lhs) flat.simps(6)} & $\dn$ & @{thm (rhs) flat.simps(6)}\\
+  @{thm (lhs) flat.simps(7)} & $\dn$ & @{thm (rhs) flat.simps(7)}\\
+  \end{tabular}
+  \end{center}
+
+  \noindent Sulzmann and Lu also define inductively an inhabitation relation
+  that associates values to regular expressions:
+
+  \begin{center}
+  \begin{tabular}{c}
+  \\[-8mm]
+  @{thm[mode=Axiom] Prf.intros(4)} \qquad
+  @{thm[mode=Axiom] Prf.intros(5)[of "c"]}\\[4mm]
+  @{thm[mode=Rule] Prf.intros(2)[of "v\<^sub>1" "r\<^sub>1" "r\<^sub>2"]} \qquad 
+  @{thm[mode=Rule] Prf.intros(3)[of "v\<^sub>2" "r\<^sub>1" "r\<^sub>2"]}\\[4mm]
+  @{thm[mode=Rule] Prf.intros(1)[of "v\<^sub>1" "r\<^sub>1" "v\<^sub>2" "r\<^sub>2"]}\\[4mm]
+  @{thm[mode=Axiom] Prf.intros(6)[of "r"]} \qquad  
+  @{thm[mode=Rule] Prf.intros(7)[of "v" "r" "vs"]}
+  \end{tabular}
+  \end{center}
+
+  \noindent Note that no values are associated with the regular expression
+  @{term ZERO}, and that the only value associated with the regular
+  expression @{term ONE} is @{term Void}, pronounced (if one must) as @{text
+  "Void"}. It is routine to establish how values ``inhabiting'' a regular
+  expression correspond to the language of a regular expression, namely
+
+  \begin{proposition}
+  @{thm L_flat_Prf}
+  \end{proposition}
+
+  In general there is more than one value associated with a regular
+  expression. In case of POSIX matching the problem is to calculate the
+  unique value that satisfies the (informal) POSIX rules from the
+  Introduction. Graphically the POSIX value calculation algorithm by
+  Sulzmann and Lu can be illustrated by the picture in Figure~\ref{Sulz}
+  where the path from the left to the right involving @{term derivatives}/@{const
+  nullable} is the first phase of the algorithm (calculating successive
+  \Brz's derivatives) and @{const mkeps}/@{text inj}, the path from right to
+  left, the second phase. This picture shows the steps required when a
+  regular expression, say @{text "r\<^sub>1"}, matches the string @{term
+  "[a,b,c]"}. We first build the three derivatives (according to @{term a},
+  @{term b} and @{term c}). We then use @{const nullable} to find out
+  whether the resulting derivative regular expression @{term "r\<^sub>4"}
+  can match the empty string. If yes, we call the function @{const mkeps}
+  that produces a value @{term "v\<^sub>4"} for how @{term "r\<^sub>4"} can
+  match the empty string (taking into account the POSIX constraints in case
+  there are several ways). This function is defined by the clauses:
+
+\begin{figure}[t]
+\begin{center}
+\begin{tikzpicture}[scale=2,node distance=1.3cm,
+                    every node/.style={minimum size=6mm}]
+\node (r1)  {@{term "r\<^sub>1"}};
+\node (r2) [right=of r1]{@{term "r\<^sub>2"}};
+\draw[->,line width=1mm](r1)--(r2) node[above,midway] {@{term "der a DUMMY"}};
+\node (r3) [right=of r2]{@{term "r\<^sub>3"}};
+\draw[->,line width=1mm](r2)--(r3) node[above,midway] {@{term "der b DUMMY"}};
+\node (r4) [right=of r3]{@{term "r\<^sub>4"}};
+\draw[->,line width=1mm](r3)--(r4) node[above,midway] {@{term "der c DUMMY"}};
+\draw (r4) node[anchor=west] {\;\raisebox{3mm}{@{term nullable}}};
+\node (v4) [below=of r4]{@{term "v\<^sub>4"}};
+\draw[->,line width=1mm](r4) -- (v4);
+\node (v3) [left=of v4] {@{term "v\<^sub>3"}};
+\draw[->,line width=1mm](v4)--(v3) node[below,midway] {@{text "inj r\<^sub>3 c"}};
+\node (v2) [left=of v3]{@{term "v\<^sub>2"}};
+\draw[->,line width=1mm](v3)--(v2) node[below,midway] {@{text "inj r\<^sub>2 b"}};
+\node (v1) [left=of v2] {@{term "v\<^sub>1"}};
+\draw[->,line width=1mm](v2)--(v1) node[below,midway] {@{text "inj r\<^sub>1 a"}};
+\draw (r4) node[anchor=north west] {\;\raisebox{-8mm}{@{term "mkeps"}}};
+\end{tikzpicture}
+\end{center}
+\mbox{}\\[-13mm]
+
+\caption{The two phases of the algorithm by Sulzmann \& Lu \cite{Sulzmann2014},
+matching the string @{term "[a,b,c]"}. The first phase (the arrows from 
+left to right) is \Brz's matcher building successive derivatives. If the 
+last regular expression is @{term nullable}, then the functions of the 
+second phase are called (the top-down and right-to-left arrows): first 
+@{term mkeps} calculates a value @{term "v\<^sub>4"} witnessing
+how the empty string has been recognised by @{term "r\<^sub>4"}. After
+that the function @{term inj} ``injects back'' the characters of the string into
+the values.
+\label{Sulz}}
+\end{figure} 
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) mkeps.simps(1)} & $\dn$ & @{thm (rhs) mkeps.simps(1)}\\
+  @{thm (lhs) mkeps.simps(2)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) mkeps.simps(2)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) mkeps.simps(3)[of "r\<^sub>1" "r\<^sub>2"]} & $\dn$ & @{thm (rhs) mkeps.simps(3)[of "r\<^sub>1" "r\<^sub>2"]}\\
+  @{thm (lhs) mkeps.simps(4)} & $\dn$ & @{thm (rhs) mkeps.simps(4)}\\
+  \end{tabular}
+  \end{center}
+
+  \noindent Note that this function needs only to be partially defined,
+  namely only for regular expressions that are nullable. In case @{const
+  nullable} fails, the string @{term "[a,b,c]"} cannot be matched by @{term
+  "r\<^sub>1"} and the null value @{term "None"} is returned. Note also how this function
+  makes some subtle choices leading to a POSIX value: for example if an
+  alternative regular expression, say @{term "ALT r\<^sub>1 r\<^sub>2"}, can
+  match the empty string and furthermore @{term "r\<^sub>1"} can match the
+  empty string, then we return a @{text Left}-value. The @{text
+  Right}-value will only be returned if @{term "r\<^sub>1"} cannot match the empty
+  string.
+
+  The most interesting idea from Sulzmann and Lu \cite{Sulzmann2014} is
+  the construction of a value for how @{term "r\<^sub>1"} can match the
+  string @{term "[a,b,c]"} from the value how the last derivative, @{term
+  "r\<^sub>4"} in Fig.~\ref{Sulz}, can match the empty string. Sulzmann and
+  Lu achieve this by stepwise ``injecting back'' the characters into the
+  values thus inverting the operation of building derivatives, but on the level
+  of values. The corresponding function, called @{term inj}, takes three
+  arguments, a regular expression, a character and a value. For example in
+  the first (or right-most) @{term inj}-step in Fig.~\ref{Sulz} the regular
+  expression @{term "r\<^sub>3"}, the character @{term c} from the last
+  derivative step and @{term "v\<^sub>4"}, which is the value corresponding
+  to the derivative regular expression @{term "r\<^sub>4"}. The result is
+  the new value @{term "v\<^sub>3"}. The final result of the algorithm is
+  the value @{term "v\<^sub>1"}. The @{term inj} function is defined by recursion on regular
+  expressions and by analysing the shape of values (corresponding to 
+  the derivative regular expressions).
+  %
+  \begin{center}
+  \begin{tabular}{l@ {\hspace{5mm}}lcl}
+  (1) & @{thm (lhs) injval.simps(1)} & $\dn$ & @{thm (rhs) injval.simps(1)}\\
+  (2) & @{thm (lhs) injval.simps(2)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1"]} & $\dn$ & 
+      @{thm (rhs) injval.simps(2)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1"]}\\
+  (3) & @{thm (lhs) injval.simps(3)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]} & $\dn$ & 
+      @{thm (rhs) injval.simps(3)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]}\\
+  (4) & @{thm (lhs) injval.simps(4)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]} & $\dn$ 
+      & @{thm (rhs) injval.simps(4)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]}\\
+  (5) & @{thm (lhs) injval.simps(5)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]} & $\dn$ 
+      & @{thm (rhs) injval.simps(5)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>1" "v\<^sub>2"]}\\
+  (6) & @{thm (lhs) injval.simps(6)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]} & $\dn$ 
+      & @{thm (rhs) injval.simps(6)[of "r\<^sub>1" "r\<^sub>2" "c" "v\<^sub>2"]}\\
+  (7) & @{thm (lhs) injval.simps(7)[of "r" "c" "v" "vs"]} & $\dn$ 
+      & @{thm (rhs) injval.simps(7)[of "r" "c" "v" "vs"]}\\
+  \end{tabular}
+  \end{center}
+
+  \noindent To better understand what is going on in this definition it
+  might be instructive to look first at the three sequence cases (clauses
+  (4)--(6)). In each case we need to construct an ``injected value'' for
+  @{term "SEQ r\<^sub>1 r\<^sub>2"}. This must be a value of the form @{term
+  "Seq DUMMY DUMMY"}\,. Recall the clause of the @{text derivative}-function
+  for sequence regular expressions:
+
+  \begin{center}
+  @{thm (lhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]} $\dn$ @{thm (rhs) der.simps(5)[of c "r\<^sub>1" "r\<^sub>2"]}
+  \end{center}
+
+  \noindent Consider first the @{text "else"}-branch where the derivative is @{term
+  "SEQ (der c r\<^sub>1) r\<^sub>2"}. The corresponding value must therefore
+  be of the form @{term "Seq v\<^sub>1 v\<^sub>2"}, which matches the left-hand
+  side in clause~(4) of @{term inj}. In the @{text "if"}-branch the derivative is an
+  alternative, namely @{term "ALT (SEQ (der c r\<^sub>1) r\<^sub>2) (der c
+  r\<^sub>2)"}. This means we either have to consider a @{text Left}- or
+  @{text Right}-value. In case of the @{text Left}-value we know further it
+  must be a value for a sequence regular expression. Therefore the pattern
+  we match in the clause (5) is @{term "Left (Seq v\<^sub>1 v\<^sub>2)"},
+  while in (6) it is just @{term "Right v\<^sub>2"}. One more interesting
+  point is in the right-hand side of clause (6): since in this case the
+  regular expression @{text "r\<^sub>1"} does not ``contribute'' to
+  matching the string, that means it only matches the empty string, we need to
+  call @{const mkeps} in order to construct a value for how @{term "r\<^sub>1"}
+  can match this empty string. A similar argument applies for why we can
+  expect in the left-hand side of clause (7) that the value is of the form
+  @{term "Seq v (Stars vs)"}---the derivative of a star is @{term "SEQ (der c r)
+  (STAR r)"}. Finally, the reason for why we can ignore the second argument
+  in clause (1) of @{term inj} is that it will only ever be called in cases
+  where @{term "c=d"}, but the usual linearity restrictions in patterns do
+  not allow us to build this constraint explicitly into our function
+  definition.\footnote{Sulzmann and Lu state this clause as @{thm (lhs)
+  injval.simps(1)[of "c" "c"]} $\dn$ @{thm (rhs) injval.simps(1)[of "c"]},
+  but our deviation is harmless.}
+
+  The idea of the @{term inj}-function to ``inject'' a character, say
+  @{term c}, into a value can be made precise by the first part of the
+  following lemma, which shows that the underlying string of an injected
+  value has a prepended character @{term c}; the second part shows that the
+  underlying string of an @{const mkeps}-value is always the empty string
+  (given the regular expression is nullable since otherwise @{text mkeps}
+  might not be defined).
+
+  \begin{lemma}\mbox{}\smallskip\\\label{Prf_injval_flat}
+  \begin{tabular}{ll}
+  (1) & @{thm[mode=IfThen] Prf_injval_flat}\\
+  (2) & @{thm[mode=IfThen] mkeps_flat}
+  \end{tabular}
+  \end{lemma}
+
+  \begin{proof}
+  Both properties are by routine inductions: the first one can, for example,
+  be proved by induction over the definition of @{term derivatives}; the second by
+  an induction on @{term r}. There are no interesting cases.\qed
+  \end{proof}
+
+  Having defined the @{const mkeps} and @{text inj} function we can extend
+  \Brz's matcher so that a [lexical] value is constructed (assuming the
+  regular expression matches the string). The clauses of the Sulzmann and Lu lexer are
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) lexer.simps(1)} & $\dn$ & @{thm (rhs) lexer.simps(1)}\\
+  @{thm (lhs) lexer.simps(2)} & $\dn$ & @{text "case"} @{term "lexer (der c r) s"} @{text of}\\
+                     & & \phantom{$|$} @{term "None"}  @{text "\<Rightarrow>"} @{term None}\\
+                     & & $|$ @{term "Some v"} @{text "\<Rightarrow>"} @{term "Some (injval r c v)"}                          
+  \end{tabular}
+  \end{center}
+
+  \noindent If the regular expression does not match the string, @{const None} is
+  returned. If the regular expression \emph{does}
+  match the string, then @{const Some} value is returned. One important
+  virtue of this algorithm is that it can be implemented with ease in any
+  functional programming language and also in Isabelle/HOL. In the remaining
+  part of this section we prove that this algorithm is correct.
+
+  The well-known idea of POSIX matching is informally defined by the longest
+  match and priority rule (see Introduction); as correctly argued in \cite{Sulzmann2014}, this
+  needs formal specification. Sulzmann and Lu define an ``ordering
+  relation'' between values and argue
+  that there is a maximum value, as given by the derivative-based algorithm.
+  In contrast, we shall introduce a simple inductive definition that
+  specifies directly what a \emph{POSIX value} is, incorporating the
+  POSIX-specific choices into the side-conditions of our rules. Our
+  definition is inspired by the matching relation given by Vansummeren
+  \cite{Vansummeren2006}. The relation we define is ternary and written as
+  \mbox{@{term "s \<in> r \<rightarrow> v"}}, relating strings, regular expressions and
+  values.
+  %
+  \begin{center}
+  \begin{tabular}{c}
+  @{thm[mode=Axiom] Posix.intros(1)}@{text "P"}@{term "ONE"} \qquad
+  @{thm[mode=Axiom] Posix.intros(2)}@{text "P"}@{term "c"}\medskip\\
+  @{thm[mode=Rule] Posix.intros(3)[of "s" "r\<^sub>1" "v" "r\<^sub>2"]}@{text "P+L"}\qquad
+  @{thm[mode=Rule] Posix.intros(4)[of "s" "r\<^sub>2" "v" "r\<^sub>1"]}@{text "P+R"}\medskip\\
+  $\mprset{flushleft}
+   \inferrule
+   {@{thm (prem 1) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]} \qquad
+    @{thm (prem 2) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]} \\\\
+    @{thm (prem 3) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]}}
+   {@{thm (concl) Posix.intros(5)[of "s\<^sub>1" "r\<^sub>1" "v\<^sub>1" "s\<^sub>2" "r\<^sub>2" "v\<^sub>2"]}}$@{text "PS"}\\
+  @{thm[mode=Axiom] Posix.intros(7)}@{text "P[]"}\medskip\\
+  $\mprset{flushleft}
+   \inferrule
+   {@{thm (prem 1) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \qquad
+    @{thm (prem 2) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \qquad
+    @{thm (prem 3) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]} \\\\
+    @{thm (prem 4) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]}}
+   {@{thm (concl) Posix.intros(6)[of "s\<^sub>1" "r" "v" "s\<^sub>2" "vs"]}}$@{text "P\<star>"}
+  \end{tabular}
+  \end{center}
+
+   \noindent
+   We can prove that given a string @{term s} and regular expression @{term
+   r}, the POSIX value @{term v} is uniquely determined by @{term "s \<in> r \<rightarrow> v"}.
+
+  \begin{theorem}\mbox{}\smallskip\\\label{posixdeterm}
+  \begin{tabular}{ll}
+  @{text "(1)"} & If @{thm (prem 1) Posix1(1)} then @{thm (concl)
+  Posix1(1)} and @{thm (concl) Posix1(2)}.\\
+  @{text "(2)"} & @{thm[mode=IfThen] Posix_determ(1)[of _ _ "v" "v'"]}
+  \end{tabular}
+  \end{theorem}
+
+  \begin{proof} Both by induction on the definition of @{term "s \<in> r \<rightarrow> v"}. 
+  The second parts follows by a case analysis of @{term "s \<in> r \<rightarrow> v'"} and
+  the first part.\qed
+  \end{proof}
+
+  \noindent
+  We claim that our @{term "s \<in> r \<rightarrow> v"} relation captures the idea behind the two
+  informal POSIX rules shown in the Introduction: Consider for example the
+  rules @{text "P+L"} and @{text "P+R"} where the POSIX value for a string
+  and an alternative regular expression, that is @{term "(s, ALT r\<^sub>1 r\<^sub>2)"},
+  is specified---it is always a @{text "Left"}-value, \emph{except} when the
+  string to be matched is not in the language of @{term "r\<^sub>1"}; only then it
+  is a @{text Right}-value (see the side-condition in @{text "P+R"}).
+  Interesting is also the rule for sequence regular expressions (@{text
+  "PS"}). The first two premises state that @{term "v\<^sub>1"} and @{term "v\<^sub>2"}
+  are the POSIX values for @{term "(s\<^sub>1, r\<^sub>1)"} and @{term "(s\<^sub>2, r\<^sub>2)"}
+  respectively. Consider now the third premise and note that the POSIX value
+  of this rule should match the string \mbox{@{term "s\<^sub>1 @ s\<^sub>2"}}. According to the
+  longest match rule, we want that the @{term "s\<^sub>1"} is the longest initial
+  split of \mbox{@{term "s\<^sub>1 @ s\<^sub>2"}} such that @{term "s\<^sub>2"} is still recognised
+  by @{term "r\<^sub>2"}. Let us assume, contrary to the third premise, that there
+  \emph{exist} an @{term "s\<^sub>3"} and @{term "s\<^sub>4"} such that @{term "s\<^sub>2"}
+  can be split up into a non-empty string @{term "s\<^sub>3"} and a possibly empty
+  string @{term "s\<^sub>4"}. Moreover the longer string @{term "s\<^sub>1 @ s\<^sub>3"} can be
+  matched by @{text "r\<^sub>1"} and the shorter @{term "s\<^sub>4"} can still be
+  matched by @{term "r\<^sub>2"}. In this case @{term "s\<^sub>1"} would \emph{not} be the
+  longest initial split of \mbox{@{term "s\<^sub>1 @ s\<^sub>2"}} and therefore @{term "Seq v\<^sub>1
+  v\<^sub>2"} cannot be a POSIX value for @{term "(s\<^sub>1 @ s\<^sub>2, SEQ r\<^sub>1 r\<^sub>2)"}. 
+  The main point is that our side-condition ensures the longest 
+  match rule is satisfied.
+
+  A similar condition is imposed on the POSIX value in the @{text
+  "P\<star>"}-rule. Also there we want that @{term "s\<^sub>1"} is the longest initial
+  split of @{term "s\<^sub>1 @ s\<^sub>2"} and furthermore the corresponding value
+  @{term v} cannot be flattened to the empty string. In effect, we require
+  that in each ``iteration'' of the star, some non-empty substring needs to
+  be ``chipped'' away; only in case of the empty string we accept @{term
+  "Stars []"} as the POSIX value.
+
+  Next is the lemma that shows the function @{term "mkeps"} calculates
+  the POSIX value for the empty string and a nullable regular expression.
+
+  \begin{lemma}\label{lemmkeps}
+  @{thm[mode=IfThen] Posix_mkeps}
+  \end{lemma}
+
+  \begin{proof}
+  By routine induction on @{term r}.\qed 
+  \end{proof}
+
+  \noindent
+  The central lemma for our POSIX relation is that the @{text inj}-function
+  preserves POSIX values.
+
+  \begin{lemma}\label{Posix2}
+  @{thm[mode=IfThen] Posix_injval}
+  \end{lemma}
+
+  \begin{proof}
+  By induction on @{text r}. We explain two cases.
+
+  \begin{itemize}
+  \item[$\bullet$] Case @{term "r = ALT r\<^sub>1 r\<^sub>2"}. There are
+  two subcases, namely @{text "(a)"} \mbox{@{term "v = Left v'"}} and @{term
+  "s \<in> der c r\<^sub>1 \<rightarrow> v'"}; and @{text "(b)"} @{term "v = Right v'"}, @{term
+  "s \<notin> L (der c r\<^sub>1)"} and @{term "s \<in> der c r\<^sub>2 \<rightarrow> v'"}. In @{text "(a)"} we
+  know @{term "s \<in> der c r\<^sub>1 \<rightarrow> v'"}, from which we can infer @{term "(c # s)
+  \<in> r\<^sub>1 \<rightarrow> injval r\<^sub>1 c v'"} by induction hypothesis and hence @{term "(c #
+  s) \<in> ALT r\<^sub>1 r\<^sub>2 \<rightarrow> injval (ALT r\<^sub>1 r\<^sub>2) c (Left v')"} as needed. Similarly
+  in subcase @{text "(b)"} where, however, in addition we have to use
+  Prop.~\ref{derprop}(2) in order to infer @{term "c # s \<notin> L r\<^sub>1"} from @{term
+  "s \<notin> L (der c r\<^sub>1)"}.
+
+  \item[$\bullet$] Case @{term "r = SEQ r\<^sub>1 r\<^sub>2"}. There are three subcases:
+  
+  \begin{quote}
+  \begin{description}
+  \item[@{text "(a)"}] @{term "v = Left (Seq v\<^sub>1 v\<^sub>2)"} and @{term "nullable r\<^sub>1"} 
+  \item[@{text "(b)"}] @{term "v = Right v\<^sub>1"} and @{term "nullable r\<^sub>1"} 
+  \item[@{text "(c)"}] @{term "v = Seq v\<^sub>1 v\<^sub>2"} and @{term "\<not> nullable r\<^sub>1"} 
+  \end{description}
+  \end{quote}
+
+  \noindent For @{text "(a)"} we know @{term "s\<^sub>1 \<in> der c r\<^sub>1 \<rightarrow> v\<^sub>1"} and
+  @{term "s\<^sub>2 \<in> r\<^sub>2 \<rightarrow> v\<^sub>2"} as well as
+  %
+  \[@{term "\<not> (\<exists>s\<^sub>3 s\<^sub>4. s\<^sub>3 \<noteq> [] \<and> s\<^sub>3 @ s\<^sub>4 = s\<^sub>2 \<and> s\<^sub>1 @ s\<^sub>3 \<in> L (der c r\<^sub>1) \<and> s\<^sub>4 \<in> L r\<^sub>2)"}\]
+
+  \noindent From the latter we can infer by Prop.~\ref{derprop}(2):
+  %
+  \[@{term "\<not> (\<exists>s\<^sub>3 s\<^sub>4. s\<^sub>3 \<noteq> [] \<and> s\<^sub>3 @ s\<^sub>4 = s\<^sub>2 \<and> (c # s\<^sub>1) @ s\<^sub>3 \<in> L r\<^sub>1 \<and> s\<^sub>4 \<in> L r\<^sub>2)"}\]
+
+  \noindent We can use the induction hypothesis for @{text "r\<^sub>1"} to obtain
+  @{term "(c # s\<^sub>1) \<in> r\<^sub>1 \<rightarrow> injval r\<^sub>1 c v\<^sub>1"}. Putting this all together allows us to infer
+  @{term "((c # s\<^sub>1) @ s\<^sub>2) \<in> SEQ r\<^sub>1 r\<^sub>2 \<rightarrow> Seq (injval r\<^sub>1 c v\<^sub>1) v\<^sub>2"}. The case @{text "(c)"}
+  is similar.
+
+  For @{text "(b)"} we know @{term "s \<in> der c r\<^sub>2 \<rightarrow> v\<^sub>1"} and 
+  @{term "s\<^sub>1 @ s\<^sub>2 \<notin> L (SEQ (der c r\<^sub>1) r\<^sub>2)"}. From the former
+  we have @{term "(c # s) \<in> r\<^sub>2 \<rightarrow> (injval r\<^sub>2 c v\<^sub>1)"} by induction hypothesis
+  for @{term "r\<^sub>2"}. From the latter we can infer
+  %
+  \[@{term "\<not> (\<exists>s\<^sub>3 s\<^sub>4. s\<^sub>3 \<noteq> [] \<and> s\<^sub>3 @ s\<^sub>4 = c # s \<and> s\<^sub>3 \<in> L r\<^sub>1 \<and> s\<^sub>4 \<in> L r\<^sub>2)"}\]
+
+  \noindent By Lem.~\ref{lemmkeps} we know @{term "[] \<in> r\<^sub>1 \<rightarrow> (mkeps r\<^sub>1)"}
+  holds. Putting this all together, we can conclude with @{term "(c #
+  s) \<in> SEQ r\<^sub>1 r\<^sub>2 \<rightarrow> Seq (mkeps r\<^sub>1) (injval r\<^sub>2 c v\<^sub>1)"}, as required.
+
+  Finally suppose @{term "r = STAR r\<^sub>1"}. This case is very similar to the
+  sequence case, except that we need to also ensure that @{term "flat (injval r\<^sub>1
+  c v\<^sub>1) \<noteq> []"}. This follows from @{term "(c # s\<^sub>1)
+  \<in> r\<^sub>1 \<rightarrow> injval r\<^sub>1 c v\<^sub>1"}  (which in turn follows from @{term "s\<^sub>1 \<in> der c
+  r\<^sub>1 \<rightarrow> v\<^sub>1"} and the induction hypothesis).\qed
+  \end{itemize}
+  \end{proof}
+
+  \noindent
+  With Lem.~\ref{Posix2} in place, it is completely routine to establish
+  that the Sulzmann and Lu lexer satisfies our specification (returning
+  the null value @{term "None"} iff the string is not in the language of the regular expression,
+  and returning a unique POSIX value iff the string \emph{is} in the language):
+
+  \begin{theorem}\mbox{}\smallskip\\\label{lexercorrect}
+  \begin{tabular}{ll}
+  (1) & @{thm (lhs) lexer_correct_None} if and only if @{thm (rhs) lexer_correct_None}\\
+  (2) & @{thm (lhs) lexer_correct_Some} if and only if @{thm (rhs) lexer_correct_Some}\\
+  \end{tabular}
+  \end{theorem}
+
+  \begin{proof}
+  By induction on @{term s} using Lem.~\ref{lemmkeps} and \ref{Posix2}.\qed  
+  \end{proof}
+
+  \noindent In (2) we further know by Thm.~\ref{posixdeterm} that the
+  value returned by the lexer must be unique.   A simple corollary 
+  of our two theorems is:
+
+  \begin{corollary}\mbox{}\smallskip\\\label{lexercorrectcor}
+  \begin{tabular}{ll}
+  (1) & @{thm (lhs) lexer_correctness(2)} if and only if @{thm (rhs) lexer_correctness(2)}\\ 
+  (2) & @{thm (lhs) lexer_correctness(1)} if and only if @{thm (rhs) lexer_correctness(1)}\\
+  \end{tabular}
+  \end{corollary}
+
+  \noindent
+  This concludes our
+  correctness proof. Note that we have not changed the algorithm of
+  Sulzmann and Lu,\footnote{All deviations we introduced are
+  harmless.} but introduced our own specification for what a correct
+  result---a POSIX value---should be. A strong point in favour of
+  Sulzmann and Lu's algorithm is that it can be extended in various
+  ways.
+
+*}
+
+section {* Extensions and Optimisations*}
+
+text {*
+
+  If we are interested in tokenising a string, then we need to not just
+  split up the string into tokens, but also ``classify'' the tokens (for
+  example whether it is a keyword or an identifier). This can be done with
+  only minor modifications to the algorithm by introducing \emph{record
+  regular expressions} and \emph{record values} (for example
+  \cite{Sulzmann2014b}):
+
+  \begin{center}  
+  @{text "r :="}
+  @{text "..."} $\mid$
+  @{text "(l : r)"} \qquad\qquad
+  @{text "v :="}
+  @{text "..."} $\mid$
+  @{text "(l : v)"}
+  \end{center}
+  
+  \noindent where @{text l} is a label, say a string, @{text r} a regular
+  expression and @{text v} a value. All functions can be smoothly extended
+  to these regular expressions and values. For example \mbox{@{text "(l :
+  r)"}} is nullable iff @{term r} is, and so on. The purpose of the record
+  regular expression is to mark certain parts of a regular expression and
+  then record in the calculated value which parts of the string were matched
+  by this part. The label can then serve as classification for the tokens.
+  For this recall the regular expression @{text "(r\<^bsub>key\<^esub> + r\<^bsub>id\<^esub>)\<^sup>\<star>"} for
+  keywords and identifiers from the Introduction. With the record regular
+  expression we can form \mbox{@{text "((key : r\<^bsub>key\<^esub>) + (id : r\<^bsub>id\<^esub>))\<^sup>\<star>"}}
+  and then traverse the calculated value and only collect the underlying
+  strings in record values. With this we obtain finite sequences of pairs of
+  labels and strings, for example
+
+  \[@{text "(l\<^sub>1 : s\<^sub>1), ..., (l\<^sub>n : s\<^sub>n)"}\]
+  
+  \noindent from which tokens with classifications (keyword-token,
+  identifier-token and so on) can be extracted.
+
+  Derivatives as calculated by \Brz's method are usually more complex
+  regular expressions than the initial one; the result is that the
+  derivative-based matching and lexing algorithms are often abysmally slow.
+  However, various optimisations are possible, such as the simplifications
+  of @{term "ALT ZERO r"}, @{term "ALT r ZERO"}, @{term "SEQ ONE r"} and
+  @{term "SEQ r ONE"} to @{term r}. These simplifications can speed up the
+  algorithms considerably, as noted in \cite{Sulzmann2014}. One of the
+  advantages of having a simple specification and correctness proof is that
+  the latter can be refined to prove the correctness of such simplification
+  steps. While the simplification of regular expressions according to 
+  rules like
+
+  \begin{equation}\label{Simpl}
+  \begin{array}{lcllcllcllcl}
+  @{term "ALT ZERO r"} & @{text "\<Rightarrow>"} & @{term r} \hspace{8mm}%\\
+  @{term "ALT r ZERO"} & @{text "\<Rightarrow>"} & @{term r} \hspace{8mm}%\\
+  @{term "SEQ ONE r"}  & @{text "\<Rightarrow>"} & @{term r} \hspace{8mm}%\\
+  @{term "SEQ r ONE"}  & @{text "\<Rightarrow>"} & @{term r}
+  \end{array}
+  \end{equation}
+
+  \noindent is well understood, there is an obstacle with the POSIX value
+  calculation algorithm by Sulzmann and Lu: if we build a derivative regular
+  expression and then simplify it, we will calculate a POSIX value for this
+  simplified derivative regular expression, \emph{not} for the original (unsimplified)
+  derivative regular expression. Sulzmann and Lu \cite{Sulzmann2014} overcome this obstacle by
+  not just calculating a simplified regular expression, but also calculating
+  a \emph{rectification function} that ``repairs'' the incorrect value.
+  
+  The rectification functions can be (slightly clumsily) implemented  in
+  Isabelle/HOL as follows using some auxiliary functions:
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) F_RIGHT.simps(1)} & $\dn$ & @{text "Right (f v)"}\\
+  @{thm (lhs) F_LEFT.simps(1)} & $\dn$ & @{text "Left (f v)"}\\
+  
+  @{thm (lhs) F_ALT.simps(1)} & $\dn$ & @{text "Right (f\<^sub>2 v)"}\\
+  @{thm (lhs) F_ALT.simps(2)} & $\dn$ & @{text "Left (f\<^sub>1 v)"}\\
+  
+  @{thm (lhs) F_SEQ1.simps(1)} & $\dn$ & @{text "Seq (f\<^sub>1 ()) (f\<^sub>2 v)"}\\
+  @{thm (lhs) F_SEQ2.simps(1)} & $\dn$ & @{text "Seq (f\<^sub>1 v) (f\<^sub>2 ())"}\\
+  @{thm (lhs) F_SEQ.simps(1)} & $\dn$ & @{text "Seq (f\<^sub>1 v\<^sub>1) (f\<^sub>2 v\<^sub>2)"}\medskip\\
+  %\end{tabular}
+  %
+  %\begin{tabular}{lcl}
+  @{term "simp_ALT (ZERO, DUMMY) (r\<^sub>2, f\<^sub>2)"} & $\dn$ & @{term "(r\<^sub>2, F_RIGHT f\<^sub>2)"}\\
+  @{term "simp_ALT (r\<^sub>1, f\<^sub>1) (ZERO, DUMMY)"} & $\dn$ & @{term "(r\<^sub>1, F_LEFT f\<^sub>1)"}\\
+  @{term "simp_ALT (r\<^sub>1, f\<^sub>1) (r\<^sub>2, f\<^sub>2)"} & $\dn$ & @{term "(ALT r\<^sub>1 r\<^sub>2, F_ALT f\<^sub>1 f\<^sub>2)"}\\
+  @{term "simp_SEQ (ONE, f\<^sub>1) (r\<^sub>2, f\<^sub>2)"} & $\dn$ & @{term "(r\<^sub>2, F_SEQ1 f\<^sub>1 f\<^sub>2)"}\\
+  @{term "simp_SEQ (r\<^sub>1, f\<^sub>1) (ONE, f\<^sub>2)"} & $\dn$ & @{term "(r\<^sub>1, F_SEQ2 f\<^sub>1 f\<^sub>2)"}\\
+  @{term "simp_SEQ (r\<^sub>1, f\<^sub>1) (r\<^sub>2, f\<^sub>2)"} & $\dn$ & @{term "(SEQ r\<^sub>1 r\<^sub>2, F_SEQ f\<^sub>1 f\<^sub>2)"}\\
+  \end{tabular}
+  \end{center}
+
+  \noindent
+  The functions @{text "simp\<^bsub>Alt\<^esub>"} and @{text "simp\<^bsub>Seq\<^esub>"} encode the simplification rules
+  in \eqref{Simpl} and compose the rectification functions (simplifications can occur
+  deep inside the regular expression). The main simplification function is then 
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{term "simp (ALT r\<^sub>1 r\<^sub>2)"} & $\dn$ & @{term "simp_ALT (simp r\<^sub>1) (simp r\<^sub>2)"}\\
+  @{term "simp (SEQ r\<^sub>1 r\<^sub>2)"} & $\dn$ & @{term "simp_SEQ (simp r\<^sub>1) (simp r\<^sub>2)"}\\
+  @{term "simp r"} & $\dn$ & @{term "(r, id)"}\\
+  \end{tabular}
+  \end{center} 
+
+  \noindent where @{term "id"} stands for the identity function. The
+  function @{const simp} returns a simplified regular expression and a corresponding
+  rectification function. Note that we do not simplify under stars: this
+  seems to slow down the algorithm, rather than speed it up. The optimised
+  lexer is then given by the clauses:
+  
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{thm (lhs) slexer.simps(1)} & $\dn$ & @{thm (rhs) slexer.simps(1)}\\
+  @{thm (lhs) slexer.simps(2)} & $\dn$ & 
+                         @{text "let (r\<^sub>s, f\<^sub>r) = simp (r "}$\backslash$@{text " c) in"}\\
+                     & & @{text "case"} @{term "slexer r\<^sub>s s"} @{text of}\\
+                     & & \phantom{$|$} @{term "None"}  @{text "\<Rightarrow>"} @{term None}\\
+                     & & $|$ @{term "Some v"} @{text "\<Rightarrow>"} @{text "Some (inj r c (f\<^sub>r v))"}                          
+  \end{tabular}
+  \end{center}
+
+  \noindent
+  In the second clause we first calculate the derivative @{term "der c r"}
+  and then simplify the result. This gives us a simplified derivative
+  @{text "r\<^sub>s"} and a rectification function @{text "f\<^sub>r"}. The lexer
+  is then recursively called with the simplified derivative, but before
+  we inject the character @{term c} into the value @{term v}, we need to rectify
+  @{term v} (that is construct @{term "f\<^sub>r v"}). Before we can establish the correctness
+  of @{term "slexer"}, we need to show that simplification preserves the language
+  and simplification preserves our POSIX relation once the value is rectified
+  (recall @{const "simp"} generates a (regular expression, rectification function) pair):
+
+  \begin{lemma}\mbox{}\smallskip\\\label{slexeraux}
+  \begin{tabular}{ll}
+  (1) & @{thm L_fst_simp[symmetric]}\\
+  (2) & @{thm[mode=IfThen] Posix_simp}
+  \end{tabular}
+  \end{lemma}
+
+  \begin{proof} Both are by induction on @{text r}. There is no
+  interesting case for the first statement. For the second statement,
+  of interest are the @{term "r = ALT r\<^sub>1 r\<^sub>2"} and @{term "r = SEQ r\<^sub>1
+  r\<^sub>2"} cases. In each case we have to analyse four subcases whether
+  @{term "fst (simp r\<^sub>1)"} and @{term "fst (simp r\<^sub>2)"} equals @{const
+  ZERO} (respectively @{const ONE}). For example for @{term "r = ALT
+  r\<^sub>1 r\<^sub>2"}, consider the subcase @{term "fst (simp r\<^sub>1) = ZERO"} and
+  @{term "fst (simp r\<^sub>2) \<noteq> ZERO"}. By assumption we know @{term "s \<in>
+  fst (simp (ALT r\<^sub>1 r\<^sub>2)) \<rightarrow> v"}. From this we can infer @{term "s \<in> fst (simp r\<^sub>2) \<rightarrow> v"}
+  and by IH also (*) @{term "s \<in> r\<^sub>2 \<rightarrow> (snd (simp r\<^sub>2) v)"}. Given @{term "fst (simp r\<^sub>1) = ZERO"}
+  we know @{term "L (fst (simp r\<^sub>1)) = {}"}. By the first statement
+  @{term "L r\<^sub>1"} is the empty set, meaning (**) @{term "s \<notin> L r\<^sub>1"}.
+  Taking (*) and (**) together gives by the \mbox{@{text "P+R"}}-rule 
+  @{term "s \<in> ALT r\<^sub>1 r\<^sub>2 \<rightarrow> Right (snd (simp r\<^sub>2) v)"}. In turn this
+  gives @{term "s \<in> ALT r\<^sub>1 r\<^sub>2 \<rightarrow> snd (simp (ALT r\<^sub>1 r\<^sub>2)) v"} as we need to show.
+  The other cases are similar.\qed
+  \end{proof}
+
+  \noindent We can now prove relatively straightforwardly that the
+  optimised lexer produces the expected result:
+
+  \begin{theorem}
+  @{thm slexer_correctness}
+  \end{theorem}
+
+  \begin{proof} By induction on @{term s} generalising over @{term
+  r}. The case @{term "[]"} is trivial. For the cons-case suppose the
+  string is of the form @{term "c # s"}. By induction hypothesis we
+  know @{term "slexer r s = lexer r s"} holds for all @{term r} (in
+  particular for @{term "r"} being the derivative @{term "der c
+  r"}). Let @{term "r\<^sub>s"} be the simplified derivative regular expression, that is @{term
+  "fst (simp (der c r))"}, and @{term "f\<^sub>r"} be the rectification
+  function, that is @{term "snd (simp (der c r))"}.  We distinguish the cases
+  whether (*) @{term "s \<in> L (der c r)"} or not. In the first case we
+  have by Thm.~\ref{lexercorrect}(2) a value @{term "v"} so that @{term
+  "lexer (der c r) s = Some v"} and @{term "s \<in> der c r \<rightarrow> v"} hold.
+  By Lem.~\ref{slexeraux}(1) we can also infer from~(*) that @{term "s
+  \<in> L r\<^sub>s"} holds.  Hence we know by Thm.~\ref{lexercorrect}(2) that
+  there exists a @{term "v'"} with @{term "lexer r\<^sub>s s = Some v'"} and
+  @{term "s \<in> r\<^sub>s \<rightarrow> v'"}. From the latter we know by
+  Lem.~\ref{slexeraux}(2) that @{term "s \<in> der c r \<rightarrow> (f\<^sub>r v')"} holds.
+  By the uniqueness of the POSIX relation (Thm.~\ref{posixdeterm}) we
+  can infer that @{term v} is equal to @{term "f\<^sub>r v'"}---that is the 
+  rectification function applied to @{term "v'"}
+  produces the original @{term "v"}.  Now the case follows by the
+  definitions of @{const lexer} and @{const slexer}.
+
+  In the second case where @{term "s \<notin> L (der c r)"} we have that
+  @{term "lexer (der c r) s = None"} by Thm.~\ref{lexercorrect}(1).  We
+  also know by Lem.~\ref{slexeraux}(1) that @{term "s \<notin> L r\<^sub>s"}. Hence
+  @{term "lexer r\<^sub>s s = None"} by Thm.~\ref{lexercorrect}(1) and
+  by IH then also @{term "slexer r\<^sub>s s = None"}. With this we can
+  conclude in this case too.\qed   
+
+  \end{proof} 
+*}
+
+section {* The Correctness Argument by Sulzmann and Lu\label{argu} *}
+
+text {*
+%  \newcommand{\greedy}{\succcurlyeq_{gr}}
+ \newcommand{\posix}{>}
+
+  An extended version of \cite{Sulzmann2014} is available at the website of
+  its first author; this includes some ``proofs'', claimed in
+  \cite{Sulzmann2014} to be ``rigorous''. Since these are evidently not in
+  final form, we make no comment thereon, preferring to give general reasons
+  for our belief that the approach of \cite{Sulzmann2014} is problematic.
+  Their central definition is an ``ordering relation'' defined by the
+  rules (slightly adapted to fit our notation):
+
+\begin{center}  
+\begin{tabular}{@ {}c@ {\hspace{4mm}}c@ {}}	
+@{thm[mode=Rule] C2[of "v\<^sub>1" "r\<^sub>1" "v\<^sub>1\<iota>" "v\<^sub>2" "r\<^sub>2" "v\<^sub>2\<iota>"]}\,(C2) &
+@{thm[mode=Rule] C1[of "v\<^sub>2" "r\<^sub>2" "v\<^sub>2\<iota>" "v\<^sub>1" "r\<^sub>1"]}\,(C1)\smallskip\\
+
+@{thm[mode=Rule] A1[of "v\<^sub>1" "v\<^sub>2" "r\<^sub>1" "r\<^sub>2"]}\,(A1) &
+@{thm[mode=Rule] A2[of "v\<^sub>2" "v\<^sub>1" "r\<^sub>1" "r\<^sub>2"]}\,(A2)\smallskip\\
+
+@{thm[mode=Rule] A3[of "v\<^sub>1" "r\<^sub>2" "v\<^sub>2" "r\<^sub>1"]}\,(A3) &
+@{thm[mode=Rule] A4[of "v\<^sub>1" "r\<^sub>1" "v\<^sub>2" "r\<^sub>2"]}\,(A4)\smallskip\\
+
+@{thm[mode=Rule] K1[of "v" "vs" "r"]}\,(K1) &
+@{thm[mode=Rule] K2[of "v" "vs" "r"]}\,(K2)\smallskip\\
+
+@{thm[mode=Rule] K3[of "v\<^sub>1" "r" "v\<^sub>2" "vs\<^sub>1" "vs\<^sub>2"]}\,(K3) &
+@{thm[mode=Rule] K4[of "vs\<^sub>1" "r" "vs\<^sub>2" "v"]}\,(K4)
+\end{tabular}
+\end{center}
+
+  \noindent The idea behind the rules (A1) and (A2), for example, is that a
+  @{text Left}-value is bigger than a @{text Right}-value, if the underlying
+  string of the @{text Left}-value is longer or of equal length to the
+  underlying string of the @{text Right}-value. The order is reversed,
+  however, if the @{text Right}-value can match a longer string than a
+  @{text Left}-value. In this way the POSIX value is supposed to be the
+  biggest value for a given string and regular expression.
+
+  Sulzmann and Lu explicitly refer to the paper \cite{Frisch2004} by Frisch
+  and Cardelli from where they have taken the idea for their correctness
+  proof. Frisch and Cardelli introduced a similar ordering for GREEDY
+  matching and they showed that their GREEDY matching algorithm always
+  produces a maximal element according to this ordering (from all possible
+  solutions). The only difference between their GREEDY ordering and the
+  ``ordering'' by Sulzmann and Lu is that GREEDY always prefers a @{text
+  Left}-value over a @{text Right}-value, no matter what the underlying
+  string is. This seems to be only a very minor difference, but it has
+  drastic consequences in terms of what properties both orderings enjoy.
+  What is interesting for our purposes is that the properties reflexivity,
+  totality and transitivity for this GREEDY ordering can be proved
+  relatively easily by induction.
+
+  These properties of GREEDY, however, do not transfer to the POSIX
+  ``ordering'' by Sulzmann and Lu, which they define as @{text "v\<^sub>1 \<ge>\<^sub>r v\<^sub>2"}. 
+  To start with, @{text "v\<^sub>1 \<ge>\<^sub>r v\<^sub>2"} is
+  not defined inductively, but as $($@{term "v\<^sub>1 = v\<^sub>2"}$)$ $\vee$ $($@{term "(v\<^sub>1 >r
+  v\<^sub>2) \<and> (flat v\<^sub>1 = flat (v\<^sub>2::val))"}$)$. This means that @{term "v\<^sub>1
+  >(r::rexp) (v\<^sub>2::val)"} does not necessarily imply @{term "v\<^sub>1 \<ge>(r::rexp)
+  (v\<^sub>2::val)"}. Moreover, transitivity does not hold in the ``usual''
+  formulation, for example:
+
+  \begin{falsehood}
+  Suppose @{term "\<turnstile> v\<^sub>1 : r"}, @{term "\<turnstile> v\<^sub>2 : r"} and @{term "\<turnstile> v\<^sub>3 : r"}.
+  If @{term "v\<^sub>1 >(r::rexp) (v\<^sub>2::val)"} and @{term "v\<^sub>2 >(r::rexp) (v\<^sub>3::val)"}
+  then @{term "v\<^sub>1 >(r::rexp) (v\<^sub>3::val)"}.
+  \end{falsehood}
+
+  \noindent If formulated in this way, then there are various counter
+  examples: For example let @{term r} be @{text "a + ((a + a)\<cdot>(a + \<zero>))"}
+  then the @{term "v\<^sub>1"}, @{term "v\<^sub>2"} and @{term "v\<^sub>3"} below are values
+  of @{term r}:
+
+  \begin{center}
+  \begin{tabular}{lcl}
+  @{term "v\<^sub>1"} & $=$ & @{term "Left(Char a)"}\\
+  @{term "v\<^sub>2"} & $=$ & @{term "Right(Seq (Left(Char a)) (Right Void))"}\\
+  @{term "v\<^sub>3"} & $=$ & @{term "Right(Seq (Right(Char a)) (Left(Char a)))"}
+  \end{tabular}
+  \end{center}
+
+  \noindent Moreover @{term "v\<^sub>1 >(r::rexp) v\<^sub>2"} and @{term "v\<^sub>2 >(r::rexp)
+  v\<^sub>3"}, but \emph{not} @{term "v\<^sub>1 >(r::rexp) v\<^sub>3"}! The reason is that
+  although @{term "v\<^sub>3"} is a @{text "Right"}-value, it can match a longer
+  string, namely @{term "flat v\<^sub>3 = [a,a]"}, while @{term "flat v\<^sub>1"} (and
+  @{term "flat v\<^sub>2"}) matches only @{term "[a]"}. So transitivity in this
+  formulation does not hold---in this example actually @{term "v\<^sub>3
+  >(r::rexp) v\<^sub>1"}!
+
+  Sulzmann and Lu ``fix'' this problem by weakening the transitivity
+  property. They require in addition that the underlying strings are of the
+  same length. This excludes the counter example above and any
+  counter-example we were able to find (by hand and by machine). Thus the
+  transitivity lemma should be formulated as:
+
+  \begin{conject}
+  Suppose @{term "\<turnstile> v\<^sub>1 : r"}, @{term "\<turnstile> v\<^sub>2 : r"} and @{term "\<turnstile> v\<^sub>3 : r"},
+  and also @{text "|v\<^sub>1| = |v\<^sub>2| = |v\<^sub>3|"}.\\
+  If @{term "v\<^sub>1 >(r::rexp) (v\<^sub>2::val)"} and @{term "v\<^sub>2 >(r::rexp) (v\<^sub>3::val)"}
+  then @{term "v\<^sub>1 >(r::rexp) (v\<^sub>3::val)"}.
+  \end{conject}
+
+  \noindent While we agree with Sulzmann and Lu that this property
+  probably(!) holds, proving it seems not so straightforward: although one
+  begins with the assumption that the values have the same flattening, this
+  cannot be maintained as one descends into the induction. This is a problem
+  that occurs in a number of places in the proofs by Sulzmann and Lu.
+
+  Although they do not give an explicit proof of the transitivity property,
+  they give a closely related property about the existence of maximal
+  elements. They state that this can be verified by an induction on @{term r}. We
+  disagree with this as we shall show next in case of transitivity. The case
+  where the reasoning breaks down is the sequence case, say @{term "SEQ r\<^sub>1 r\<^sub>2"}.
+  The induction hypotheses in this case are
+
+\begin{center}
+\begin{tabular}{@ {}c@ {\hspace{10mm}}c@ {}}
+\begin{tabular}{@ {}l@ {\hspace{-7mm}}l@ {}}
+IH @{term "r\<^sub>1"}:\\
+@{text "\<forall> v\<^sub>1, v\<^sub>2, v\<^sub>3."} \\
+  & @{term "\<turnstile> v\<^sub>1 : r\<^sub>1"}\;@{text "\<and>"}
+    @{term "\<turnstile> v\<^sub>2 : r\<^sub>1"}\;@{text "\<and>"}
+    @{term "\<turnstile> v\<^sub>3 : r\<^sub>1"}\\
+  & @{text "\<and>"} @{text "|v\<^sub>1| = |v\<^sub>2| = |v\<^sub>3|"}\\
+  & @{text "\<and>"} @{term "v\<^sub>1 >(r\<^sub>1::rexp) v\<^sub>2 \<and> v\<^sub>2 >(r\<^sub>1::rexp) v\<^sub>3"}\medskip\\
+  & $\Rightarrow$ @{term "v\<^sub>1 >(r\<^sub>1::rexp) v\<^sub>3"}
+\end{tabular} &
+\begin{tabular}{@ {}l@ {\hspace{-7mm}}l@ {}}
+IH @{term "r\<^sub>2"}:\\
+@{text "\<forall> v\<^sub>1, v\<^sub>2, v\<^sub>3."}\\ 
+  & @{term "\<turnstile> v\<^sub>1 : r\<^sub>2"}\;@{text "\<and>"}
+    @{term "\<turnstile> v\<^sub>2 : r\<^sub>2"}\;@{text "\<and>"}
+    @{term "\<turnstile> v\<^sub>3 : r\<^sub>2"}\\
+  & @{text "\<and>"} @{text "|v\<^sub>1| = |v\<^sub>2| = |v\<^sub>3|"}\\
+  & @{text "\<and>"} @{term "v\<^sub>1 >(r\<^sub>2::rexp) v\<^sub>2 \<and> v\<^sub>2 >(r\<^sub>2::rexp) v\<^sub>3"}\medskip\\
+  & $\Rightarrow$ @{term "v\<^sub>1 >(r\<^sub>2::rexp) v\<^sub>3"}
+\end{tabular}
+\end{tabular}
+\end{center} 
+
+  \noindent We can assume that
+  %
+  \begin{equation}
+  @{term "(Seq (v\<^sub>1\<^sub>l) (v\<^sub>1\<^sub>r)) >(SEQ r\<^sub>1 r\<^sub>2) (Seq (v\<^sub>2\<^sub>l) (v\<^sub>2\<^sub>r))"}
+  \qquad\textrm{and}\qquad
+  @{term "(Seq (v\<^sub>2\<^sub>l) (v\<^sub>2\<^sub>r)) >(SEQ r\<^sub>1 r\<^sub>2) (Seq (v\<^sub>3\<^sub>l) (v\<^sub>3\<^sub>r))"}
+  \label{assms}
+  \end{equation}
+
+  \noindent hold, and furthermore that the values have equal length, namely:
+  %
+  \begin{equation}
+  @{term "flat (Seq (v\<^sub>1\<^sub>l) (v\<^sub>1\<^sub>r)) = flat (Seq (v\<^sub>2\<^sub>l) (v\<^sub>2\<^sub>r))"}
+  \qquad\textrm{and}\qquad
+  @{term "flat (Seq (v\<^sub>2\<^sub>l) (v\<^sub>2\<^sub>r)) = flat (Seq (v\<^sub>3\<^sub>l) (v\<^sub>3\<^sub>r))"}
+  \label{lens}
+  \end{equation} 
+
+  \noindent We need to show that @{term "(Seq (v\<^sub>1\<^sub>l) (v\<^sub>1\<^sub>r)) >(SEQ r\<^sub>1 r\<^sub>2)
+  (Seq (v\<^sub>3\<^sub>l) (v\<^sub>3\<^sub>r))"} holds. We can proceed by analysing how the
+  assumptions in \eqref{assms} have arisen. There are four cases. Let us
+  assume we are in the case where we know
+
+  \[
+  @{term "v\<^sub>1\<^sub>l >(r\<^sub>1::rexp) v\<^sub>2\<^sub>l"}
+  \qquad\textrm{and}\qquad
+  @{term "v\<^sub>2\<^sub>l >(r\<^sub>1::rexp) v\<^sub>3\<^sub>l"}
+  \]
+
+  \noindent and also know the corresponding inhabitation judgements. This is
+  exactly a case where we would like to apply the induction hypothesis
+  IH~$r_1$. But we cannot! We still need to show that @{term "flat (v\<^sub>1\<^sub>l) =
+  flat(v\<^sub>2\<^sub>l)"} and @{term "flat(v\<^sub>2\<^sub>l) = flat(v\<^sub>3\<^sub>l)"}. We know from
+  \eqref{lens} that the lengths of the sequence values are equal, but from
+  this we cannot infer anything about the lengths of the component values.
+  Indeed in general they will be unequal, that is
+
+  \[
+  @{term "flat(v\<^sub>1\<^sub>l) \<noteq> flat(v\<^sub>2\<^sub>l)"}  
+  \qquad\textrm{and}\qquad
+  @{term "flat(v\<^sub>1\<^sub>r) \<noteq> flat(v\<^sub>2\<^sub>r)"}
+  \]
+
+  \noindent but still \eqref{lens} will hold. Now we are stuck, since the IH
+  does not apply. As said, this problem where the induction hypothesis does
+  not apply arises in several places in the proof of Sulzmann and Lu, not
+  just for proving transitivity.
+
+*}
+
+section {* Conclusion *}
+
+text {*
+
+  We have implemented the POSIX value calculation algorithm introduced by
+  Sulzmann and Lu
+  \cite{Sulzmann2014}. Our implementation is nearly identical to the
+  original and all modifications we introduced are harmless (like our char-clause for
+  @{text inj}). We have proved this algorithm to be correct, but correct
+  according to our own specification of what POSIX values are. Our
+  specification (inspired from work by Vansummeren \cite{Vansummeren2006}) appears to be
+  much simpler than in \cite{Sulzmann2014} and our proofs are nearly always
+  straightforward. We have attempted to formalise the original proof
+  by Sulzmann and Lu \cite{Sulzmann2014}, but we believe it contains
+  unfillable gaps. In the online version of \cite{Sulzmann2014}, the authors
+  already acknowledge some small problems, but our experience suggests
+  that there are more serious problems. 
+  
+  Having proved the correctness of the POSIX lexing algorithm in
+  \cite{Sulzmann2014}, which lessons have we learned? Well, this is a
+  perfect example for the importance of the \emph{right} definitions. We
+  have (on and off) explored mechanisations as soon as first versions
+  of \cite{Sulzmann2014} appeared, but have made little progress with
+  turning the relatively detailed proof sketch in \cite{Sulzmann2014} into a
+  formalisable proof. Having seen \cite{Vansummeren2006} and adapted the
+  POSIX definition given there for the algorithm by Sulzmann and Lu made all
+  the difference: the proofs, as said, are nearly straightforward. The
+  question remains whether the original proof idea of \cite{Sulzmann2014},
+  potentially using our result as a stepping stone, can be made to work?
+  Alas, we really do not know despite considerable effort.
+
+
+  Closely related to our work is an automata-based lexer formalised by
+  Nipkow \cite{Nipkow98}. This lexer also splits up strings into longest
+  initial substrings, but Nipkow's algorithm is not completely
+  computational. The algorithm by Sulzmann and Lu, in contrast, can be
+  implemented with ease in any functional language. A bespoke lexer for the
+  Imp-language is formalised in Coq as part of the Software Foundations book
+  by Pierce et al \cite{Pierce2015}. The disadvantage of such bespoke lexers is that they
+  do not generalise easily to more advanced features.
+  Our formalisation is available from the Archive of Formal Proofs \cite{aduAFP16}
+  under \url{http://www.isa-afp.org/entries/Posix-Lexing.shtml}.\medskip
+
+  \noindent
+  {\bf Acknowledgements:}
+  We are very grateful to Martin Sulzmann for his comments on our work and 
+  moreover for patiently explaining to us the details in \cite{Sulzmann2014}. We
+  also received very helpful comments from James Cheney and anonymous referees.
+  %  \small
+  \bibliographystyle{plain}
+  \bibliography{root}
+
+*}
+
+
+(*<*)
+end
+(*>*)
\ No newline at end of file