2 Let us introduce an ordering $\preceq$ between classes. Formally, \class{p}
3 $\preceq$ \class{q} if there exists a path of length $\alpha$
4 ($0<\alpha<|\mathcal{K}|$) from an element of \class{p} to an element of
5 \class{q}. One can remark that if \class{p}$\preceq$\class{q}, then it is not
6 possible to also have \class{q}$\preceq$\class{p}.
9 % The relation $ \preceq$ is a partial order
11 % Reflexivity is established since for any $p_0 \in [p]$ there exists a path
12 % of length 0 from $p_0$ to $p_0$.
13 % For antisymmetry, suppose there exists two classes $[p]$ and
14 % $[q]$ such that $[p] \preceq [q]$ and $[q] \preceq [p]$.
15 % There exists then $p_0,p'_0 \in [p]$, $q_0, q'_0 \in [q]$ with paths
16 % from $p_0$ to $q_0$ and from $q'_0$ to $p_0$.
17 % Thus, elements $p_0$, $p'_0$, $q_0$ and $q'_0$ belong to the same
18 % strongly connected component and then
19 % $[p]$ = $[q]$. Transitivity is obviously established.
24 There exists a renaming process method that assigns new identifier numbers to
25 processes $i\in$ \class{p} and $j \in$ \class{q} s.t. $i \le j$ provided
26 \class{p} $\preceq$ \class{q}.
28 % We first define a renaming process method and later show that it fulfills
31 First of all, let \class{p_1}, \ldots, \class{p_l} be classes of
32 $n_1$,\ldots, $n_l$ elements respectively that do not depend on other
33 classes. Elements of \class{p_1} are renamed by $1$, \ldots, $n_1$ and
34 elements of \class{p_i}, $2 \le i \le l$ are renamed by $1+
35 \Sigma_{k=1}^{i-1} n_k$, \ldots, $\Sigma_{k=1}^{i} n_k$. We now consider
36 the classes \class{p_1}, \ldots, \class{p_{l'}} whose elements have been
37 renamed and let $m$ be the maximum index of elements of \class{p_1}, \ldots,
38 \class{p_{l'}}. Given another class \class{p} that exclusively depends on
39 some \class{p_i}, $1 \le i \le l'$ and that contains $k$ elements. Elements
40 of \class{p} are then renamed by $m+1$, \ldots, $m+k$.
41 % In the end for remaining classes, i.e. those which do not depend on anything,
42 % elements are arbitrarily numbered.
43 This renaming process method has then been applied on $l'+1$ classes. It
44 ends since it decreases the number of elements to assign a number. Notice
45 that this process is not determinist.
47 We are then ready to prove that this renaming method verifies property
48 expressed in the lemma. The proof is done by induction on the length $l$ of
49 the longest dependency path between classes.
51 First, if \class{p} $\preceq$ \class{q} and \class{q} immediately depends on
52 \class{p}, \textit{i.e.} the longest path from elements of \class{p} to
53 elements of \class{q} has length 1. Due to the renaming method, all element
54 numbers of \class{q} are greater than the ones of \class{p} and the result
55 is established. Let \class{p} and \class{q} s.t. the longest dependency
56 path from \class{p} to \class{q} has length $l+1$. Then there exists some
57 \class{q'} s.t. \class{q} immediately depends on \class{q'} and the longest
58 dependency path from \class{p} and \class{q'} has length $l$. We have then
59 \class{q'} $\preceq$ \class{q} and then for all $k$, $j$ s.t. $k \in$
60 \class{q'} and $j \in$ \class{q}, $k \le j$. By induction hypothesis
61 \class{p} $\preceq$ \class{q'} and then for all $i$, $k$ s.t. $i \in$
62 \class{p} and $k \in$ \class{q'}, $i \le k$ and the result is established.
66 It can be noticed that the renaming process is inspired from \emph{graphs by
67 layer} of Golès and Salinas~\cite{GS08}. It ensures that identifier numbers of
68 a layer are greater than all the identifier numbers of any lower layers.
71 We have \class{1} $=\{1,2\}$, \class{3} $=\{3\}$ and \class{4} $=\{4,5\}$.
72 Processes numbers are already compliant with the order $\preceq$.
75 \begin{Proof}[of Theorem~\ref{th:cvg}]
77 % Since $\preceq$ is a partial order, [[JFC citer un theoreme qui dit cela ou
78 % le prouver]] there exists a renaming process that assigns new identifier
79 % number to processes $i\in [p]$ and $j \in [q]$ s.t. $i \le j$ provided $[p]
81 % % $i \in [p]$ and $j \in [q]$.
83 The rest of the proof is done by induction on the class index. Let us
84 consider the first class \class{b_1} with $n_1$ elements \textit{i.e.} the
85 class with the smallest identifiers.
87 By theorem hypotheses, %following the strategy $(S^t)$,
88 synchronous iterations converge to the fixed point in a finite number of
89 iterations. %pseudo periods. % [[JFC : borner m1/n1]].
90 So, all \emph{source classes} (independant from any other class) will also
91 converge in mixed mode. Let us suppose now that mixed iterations with uniform
92 delays converge for classes \class{b_1}, \ldots, \class{b_k} in a time $t_k$.
93 By construction, \class{b_{k+1}} only depends on some \class{b_1}, \ldots,
94 \class{b_k} and \class{b_{k+1}}. There exists then a sufficiently large time
95 $t_0$ s.t. $D^{t_0}_{p_{k+1}p_j}$ is greater or equal to $t_k$ for any
96 $p_{k+1} \in$ \class{b_{k+1}} and $p_j \in$ \class{b_j}, $1 \le j \le k$.
98 We are then left with synchronous iterations of elements of \class{b_{k+1}}
99 starting with configurations where all the elements of \class{b_j}, $1 \le j
100 \le k$, have constant values. By theorem hypotheses, it converges.