Let {:U"} be a converging subsequence, {(u"', urn)} be the corr
esponding sequence of dual variables of SDI". We suppose that proper subsequence s have been chosen so that {(u m, v')} is also convergent (Lemma 6) and not e .2"~' and (u ~, u '~) as the limit points. The proof follows the steps of O'Neill. Consider some proposal (xT', (7~'~")~si~)) generated at cycle m by the subproblem SD~. We have for all q/> m + 1, [ q -1 -m which simply expresses the optimality of the subproblem SDqp(~ (if thi s relation were not true, it could be shown using the reasoning of Lemma 1 tha t the activity level of that proposal in SD%,) should be increased). This condition holds for all nodes differing from the terminal ones (i ~ L,,(1)). The following re lations derive from Lemma 4. We have for i~ L,,(1) and at cycle m: c, y u2)x?= * " - 6 (u,, [C,) a~A(i) When i~ L,(1), but is not equal to l, the relation becomes: 9 t;~| /31 t?l =~s [(u,,,u,D(,~)lc]+ 2 ",. s~S(i) rtl trl m ~ m x, =x,, /z, =h,, soS(l), x,- ........ =Xs~.~, s~S(1), t?l , ,'71 171 /Zs =/l s ~,p(~.). Lemma 6. {u m, v m } belongs to a compact set and there exists a subseq uence converging at(u ~-, ,, ). Proof. Because v~'(e,s for e,~:c B we have, using Theorem 24.5 of Rockafellar [ 1970], ~v?(O) ~ ~vT(o) + n~ where 7? is a positive number. B is the ball previously defined. ~v,~(0) is a compact set because 0 is in the interior of the effective domain of the function which proves the theorem. We ca n now state the convergence theorem. Theorem. Any cluster point of the sequence generated by the algorithm is an optimal solution to the problem. ci- uqi x~"+ 2 ~"_ r (U~Di~)~X~ h, <~v q, q>~m+l, ac (i) s~S(i) a~A(i) (4) The dual of D is obtained by modifying the x which is equal to the expression appearing in the profitability criter ion. Because the directional derivative is positive, increasing him from 0 to some small positive value will lend to an increase in the objective function val ue. This will be afortiori true if we let h,, vary freely in the optimization. Nested decomposition provides a feasible solution at each cycle of pha se 2, provided we keep all proposals of each subproblem in each cycle. The procedure for constructing this solution is an immediate generalization of the on e given in Ho (1974) or O'Neill (1976); it can be stated as follows. Let $'--(~7', i~ T) be the solution at cycle m. We consider at some node s, the sequence of pr oposals In order to introduce our formulation of the dual n-stage stochastic p rogramming problem, we consider the following schematic economic planning problem. We consider a multisector economy, which in each period can adapt the te chnology of its new capital stock and its activity levels. Let ~, designate a ve ctor, which is on a one to one correspondence with the technology structure of new equipment in period t and u~ be the vector of activity levels. Let ut = (u, .... , u,) and ~'t = (sr~, 9 .., ~',) be the vectors relative to the decisions up to stage t. We consider the fo llowing economic problem (EP): Problem EPIt is easy to see that each constraint i of this program (1) involves xi and all vector variables x,~, d c D(i) and requires them to belong to a conv ex set. We further assume that each C~ is a compact set which does not derive from dua lizing EP. As in O'Neill, the assumption is mainly introduced for stating the conver gence proof. This assumption, although restrictive in some problems, can be natural in others. The numerical examples given in Section 5 are general equilibrium mode ls where only relative price matters. It is possible, in general, to restate t hese problems in such a way that the relative prices belong to compact sets. Needless to say the same need not be true with partial equilibrium models. " hi'" m' constructed at cycle m'= 1,..., m and designate by X~' the vector (x~,..., xT). Let A ~ be the matrix "~ " (h~,... - .... , h~ ), dlmensloned m x rn (A~ is an upper triangular matrix). The construction of a feasible (see Lemma 2) solution at cyc le m is then stated as follows: -- The last relation holds for i = l: c,x? + E "'' " s~S(l) ~c-S(I)