DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS ARCHIS GHATE Abstract. Finite-dimensional monotropic programs form a class of convex optimization problems that includes linear programs, convex minimum cost flow problems on networks and hypernetworks, and separable convex programs with linear constrains. Countably infinite monotropic programs arise, for example, in infinite-horizon sequential decision problems and in robust optimization. Their applications encompass (i) countably infinite linear programs such as the shortest path formulations of infinite-horizon non-stationary as well as countable-state Markov decision processes; and (ii) convex minimum cost flow problems on countably infinite networks and hypernetworks. Duality results for finite-dimensional monotropic programs are as powerful as those available for linear programs. On the contrary, applicable duality results for countably infinite monotropic programs are currently non-existent owing to several mathematical pathologies in infinite dimensional sequence spaces. This paper overcomes this hurdle by first embedding the dual variables in a sequence space where the Lagrangian function is well-defined and finite. Weak duality and complementary slackness are derived using finite-dimensional proof techniques. Conditions under which zero duality gaps and strong duality between a sequence of finite-dimensional primal-dual projections of the infinite dimensional problem are preserved in the limit are established. Essentially all known duality results about countably infinite mathematical programs are recovered as special cases. Key words. convex optimization, infinite dimensional optimization, duality AMS subject classifications. 90C25, 90C46, 90C48, 90C35 1. Introduction. Finite-dimensional monotropic programs form a class of convex optimization problems wherein duality results are as powerful as those available for linear programs. Research on monotropic programs can be traced back to the seminal works of Minty [16] and Rockafellar [19, 20, 21, 22]. Indeed, in the preface to his classic book [22] on the subject, Rockafellar commented that duality in monotropic programs is as important a tool in computation as it is in theory and interpretation. A finite-dimensional monotropic program is written as n (1) min c j (x j ) (2) x S, (3) x j X j, j = 1, 2,..., n, where n is a positive integer; x is a vector in R n with real-valued components x 1, x 2,..., x n ; X j is a nonempty interval of R for each j; c j ( ) is a convex function for each j; and S is a subspace of R n (see Section 9.7 in [5]). This class of monotropic programs includes linear programs, minimum cost network flow problems, and separable convex programs with linear constraints. Rockafellar [21] was the first to prove a zero duality gap result for this class of monotropic programs. He used a variant of the ɛ-descent method for this purpose. More recently, Bertsekas [6] generalized this result to the so-called extended monotropic programs where components x j themselves could be finite-dimensional vectors. There has been a surge of interest in deriving duality results for countably infinite mathematical programs over the last two decades. Three of these studies [9, 24, 25] Submitted to the editors July 12, 2016. Funding: Funded in part by the National Science Foundation through grant #CMMI 1561918. Department of Industrial and Systems Engineering, University of Washington, Seattle, WA (archis@uw.edu). 1

2 A. GHATE 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 focused on countably infinite linear programs (CILPs) these are linear optimization problems with a countably infinite number of variables and a countably infinite number of constraints. One study focused on linear minimum cost flow problems on countably infinite networks [29], which are a special case of CILPs. The most recent ones considered convex minimum cost flow problems on countably infinite networks and hypernetworks [17]. These duality results have been applied mainly to infinitehorizon planning problems such as Markov decision processes (MDPs) [10, 11, 14] but also to robust optimization [9]. This paper extends these duality results to countably infinite monotropic programs. There is only one existing strong duality result for countably infinite monotropic programs [7]. There, the primal variables were embedded in a countable product of locally convex spaces and the dual variables resided in the topological dual of this primal space. A strong duality result was then derived under a constraint qualification that called for a certain set to be closed (see their Theorem 3.5 and Corollary 3.1). Unfortunately, such closedness conditions, although theoretically elegant, have long been thought to be hard to apply in practice (see the discussions in [2, 24, 25] for the special case of CILPs and [15] for the more general case of convex programs). Moreover, in most applications in Operations Research and Economics (including in CILPs, in infinite dimensional convex network flow problems, and in convex infinite-horizon planning problems), this choice of variable spaces would amount to selecting the space of all real sequences for the primal variables and the space of all real sequences with finitely many nonzero entries for the dual variables (this dual variable space is called the generalized finite sequence space (GFSS) [2]). This choice is unsatisfactory. For instance, in countable state MDPs, where each dual variable corresponds to the value of an MDP state, choosing GFSS as the dual variable space would imply that only a finite number of states have nonzero values. Thus, an alternative approach is needed to accommodate most applications of interest to the Operations Research and Economics communities. Since countably infinite monotropic problems are a special case of infinite dimensional convex programs, it is perhaps tempting to use Slater s constraint qualification to derive a zero duality gap result (see Theorem 3.11.2 in Ponstein [18], for instance). This would require that the positive cone of the primal constraint space has a nonempty interior. For many problems of interest in Operations Research and Economics, one natural choice for the constraint space is again the space of all real sequences. Unfortunately, the interior of the positive cone of this space is empty in its natural product topology, rendering Slater s constraint qualification inapplicable. A possible remedy for some problems is to instead use the space of all bounded sequences with its usual supremum norm topology because the positive cone of this constraint space does have a non-empty interior. Unfortunately, however, the corresponding dual problem is notoriously difficult to characterize owing to the existence of the so-called singular functionals. For instance, in CILPs, this dual problem cannot, in general, be written using the ordinary transpose of the primal constraint matrix and then the dual variables cannot be interpreted as shadow prices. These hurdles in establishing a zero duality gap result in infinite dimensional convex programs were recently formalized by Martin et al. [15] via what they called the Slater conundrum. Specifically, they made a remarkable observation by employing the notion of a core point. A core point is an algebraic counterpart of the topological concept of an interior point. Informally, they commented, on the one hand, existence of a core point ensures a zero duality gap (a desirable property), but on the other hand, existence of a core point implies the existence of singular dual functionals (an undesirable property). All in all, such

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 3 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 negative results have historically posed significant stumbling blocks in establishing concrete and applicable duality results for countably infinite mathematical programs using closedness or interior point-type sufficient conditions. As in [9, 17], these challenges are circumvented in this paper by applying a hybrid approach that (i) provides a way to choose a space for dual variables such that weak duality and complementary slackness hold, and (ii) switches to the planning horizon method to derive conditions under which zero duality gaps and strong duality in finite-dimensional approximations are preserved in the limit. The paper is organized as follows. A countably infinite monotropic program is formulated in the next section. Hypothesis (H1) on the associated convex functions is then introduced. Under H1, the objective function in the monotropic program is well-defined and finite. Two other hypotheses (H2, H3) are then stated. These define a linear subspace (Lemma 2) where the dual variables reside. H1, H2, and H3 enable a proof that the Lagrangian function is well-defined and finite (Lemma 3). This problem setup from Section 2 facilitates the use of finite-dimensional proof techniques to establish weak duality (Proposition 4) and complementary slackness (Proposition 10) in Section 3. A planning horizon approach is then devised to establish a zero duality gap result (Proposition 19) in Section 4. To achieve this, a sequence of finite-dimensional projections of the infinite dimensional monotropic program is constructed. These finite-dimensional problems are monotropic programs themselves. Optimal values of these finite-dimensional problems are shown to converge to the optimal value of the infinite dimensional problem (Proposition 18). The proof is similar to that of Berge s maximum principle and employs the concept of Kuratowski set convergence (Definition 14). The dual of the finite-dimensional monotropic program is then written, and it is shown that the infinite-dimensional dual optimal value is bounded below by the finite-dimensional dual optimal value when costs are nonnegative (Lemma 20). This, when combined with (i) a zero duality gap result in finite-dimensional monotropic programs (see Proposition 9.19 in [5]), (ii) value convergence, and finally (iii) weak duality, yields the no duality gap result. Dual value convergence is then also derived. Finally, strong duality is established in Section 5 (see Theorem 22) under the assumption that optimal dual variables in the finite-dimensional dual problems are uniformly bounded. This result does not require nonnegativity of costs. The proof derives a convergent subsequence from a sequence of pairs of optimal solutions to the finitedimensional monotropic program and its dual, and shows that its accumulation point is feasible and complementary for the infinite dimensional monotropic program and its dual. This implies, by the complementary slackness result in Section 3, that this accumulation point pair is optimal to the infinite dimensional monotropic program and its dual, and that the two problems have identical optimal values. Although, from a high-level viewpoint, the approach and hence the sequence of results in Sections 3-5 in this paper are identical to the author s previous work in [17], a more sophisticated analysis is needed to ensure that the methodology works here. For instance, the dual problems in the author s previous work did not utilize the notion of an orthogonal subspace, but this is needed to properly characterize the dual problem here. Moreover, it is harder to define finite-dimensional problems, and the idea of Kuratowski set convergence is employed in the planning horizon approach in this paper. Sections 6 and 7 show that known duality results from the author s recent work on CILPs [9] and on convex minimum cost flow problems on infinite networks and hypernetworks [17] can be recovered as special cases of the monotropic programming results here. In the finite-dimensional case, it is easy to show that linear programs and convex minimum cost flow problems, as well as the corresponding duality results, are special cases of

4 A. GHATE 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 monotropic programs. As we shall see, showing this in the countably infinite context requires considerable work. Moreover, a nontrivial argument is needed to characterize the orthogonal subspaces that define the dual problems in each case. Sections 6 and 7 therefore include some of the main contributions of this paper although this might not be evident at first glance. New duality results for countably infinite separable convex programs with linear constraints are derived in Section 8 as a special case of the monotropic programming results. 2. Problem formulation. Throughout this paper, the symbol is employed to define new notations. The symbol N will denote the set {1, 2, 3,...} of natural numbers and # will be reserved for set cardinalities. The linear vector space of all real-valued sequences x (x 1, x 2, x 3,...) is denoted by R N. The usual metrizable product topology of pointwise convergence on R N is employed throughout. Let S be any subspace of R N. For each j N, let < a j b j < + be any two real numbers, and let u j max{ a j, b j }. Suppose that, for each j N, c j ( ) is a real-valued, convex, and continuous function over the closed and bounded interval X j [a j, b j ]. Now consider the infinite dimensional optimization problem 159 160 161 162 (4) (5) (6) (P ) V inf C(x) c j (x j ) x S, x j X j, j N. 163 164 Problem (P ) will be called a countably infinite monotropic program. In the sequel, X X j will denote the Cartesian product of intervals X j across j N. The 165 feasible region of (P ) is denoted by F S X. 166 Since c j ( ) is continuous over the compact interval X j, it attains its maximum 167 on this interval, and let c j max c j (x j ). Assume that x j X j H1. the series c j of nonnegative terms is finite. 169 This hypothesis was inspired by a similar assumption that the author used in his 170 recent work on duality in CILPs [9] and in convex minimum cost flow problems in 171 countably infinite networks and hypernetworks [17]. Such hypotheses are standard in 172 the literature on infinite dimensional optimization [10, 27, 29] as they ensure that the 173 primal objective function is well-defined and finite. 174 175 176 177 178 179 180 181 182 183 184 185 Lemma 1. Suppose S is closed. If (P ) has a feasible solution, then it has an optimal solution. Proof. The series of functions c j (x j ) is easily seen to be uniformly convergent over F by applying the Weierstrass test [3] using H1. This implies that the objective function C(x) in (P ) is continuous over F (see Theorem 9.7 in [3]). The product X of compact intervals is compact by Tychonoff s theorem (see Theorem 2.61 in [1]). Hence the feasible region F is an intersection of a closed set S and a compact set X; F is thus compact (see Theorem 2.35 and its corollary in [26]). The result then follows by Corollary 2.35 in [1], which states that a continuous function achieves its minimum on a compact set. The assumption that S is closed is not vacuous because, unlike finite-dimensional Euclidean spaces, not all subspaces of R N are closed in the product topology (consider

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 5 186 187 188 189 190 191 the subspace GFSS, for instance, with the sequence of vectors whose first n components equal 1 and all tail components equal 0; this sequence converges to the vector of all ones, which is not in GFSS). We will show in Sections 6 and 7 that S is closed in many applications of interest. As in finite-dimensional monotropic programs, use a new variable y (y 1, y 2,...) to first rewrite (P ) in the equivalent format 192 (7) (P ) V inf C(x) c j (x j ) 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 (8) (9) (10) x j = y j, j N, y S, x j X j, j N. In order to write the dual problem, a dual variable λ j is attached to the equality constraint (8) for each j N. The resulting sequence λ (λ 1, λ 2, λ 3,...) needs to be embedded into an appropriate subset Λ of R N. Specifically, let Λ be the subset of all sequences λ such that H2. the series λ j u j of nonnegative terms is finite; and H3. the series λ j y j of nonnegative terms is finite for each y S. These two hypotheses were motivated by similar assumptions in the author s work on CILPs [9] and on convex minimum cost flow problems in countably infinite networks and hypernetworks [17]. H2 and H3 help ensure that two series that appear in our Lagrangian function below converge. Lemma 2. The subset Λ is a subspace of R N. Proof. Straightforward, hence omitted. Now define the Lagrangian function as (11) L(x, y; λ) c j (x j ) + λ j (y j x j ), x X, y S, λ Λ. Lemma 3. The Lagrangian is well-defined and finite for every x X, y S, and λ Λ. Proof. The first series c j (x j ) converges (absolutely) over X by H1. Also, the series λ j x j converges (absolutely) over X by H2. Similarly, the series λ j y j converges (absolutely) by H3. Thus, the series (λ j y j λ j x j ) converges by Theorem 8.8 in [3]. This means that the second series λ j (y j x j ) in the Lagrangian function also converges. This proves the claim. Now define (12) φ(λ) inf x X, y S and write the dual of (P ) as (13) (D) W = sup λ Λ L(x, y; λ), λ Λ, φ(λ).

223 224 225 226 6 A. GHATE 3. Weak duality and complementary slackness. Since c j (x j ), λ j x j and λ j y j converge by H1, H2, and H3, the Lagrangian function in (11) can be rewritten equivalently as (14) L(x, y; λ) = (c j (x j ) λ j x j ) + λ j y j, x X, y S, λ Λ. 227 228 229 230 231 232 233 234 235 236 237 238 239 240 Consequently, the function φ( ) from (12) can also be rewritten as [ ) φ(λ) = inf L(x, y; λ) = inf (c j (x j ) λ j x j + ] (15) λ j y j x X, y S x X, y S [ )] (c j (x j ) λ j x j (16) (17) = inf y S = inf y S λ j y j + inf x X λ j y j + inf {x j X j} [(c j (x j ) λ j x j )]. The function c j (x j ) λ j x j is continuous over X j for each j, and hence the second infimum above can be replaced by a minimum. Specifically, let (18) φ j (λ j ) min x j X j c j (x j ) λ j x j, j N. Then, the function φ( ) can be compactly rewritten as φ j (λ j ) if λ S, and (19) φ(λ) = otherwise, where S {λ Λ } λ j y j = 0, y S is the orthogonal complement of S. It is easy to see that S is a subspace of Λ. The dual problem (D) can then be rewritten using (19) as 241 242 243 (20) (21) (D) W = sup φ j (λ j ) λ S. 244 245 246 247 248 249 Proposition 4 (Weak duality). C(x) φ(λ) for any x F and any λ Λ. Thus, V W. Proof. The claim trivially holds by (19) if λ / S. So suppose that λ S. Then, from (19), we get, (22) φ(λ) = φ j (λ j ) ) (c j (x j ) λ j x j = c j (x j ) λ j x j = c j (x j ) = C(x). 250 251 Here, the inequality follows from (18) and the penultimate equality holds because λ S and x S. This proves the claim.

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 7 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 Corollary 5. Suppose x F and λ Λ are such that C(x ) = φ(λ ). Then x is optimal to (P ) and λ is optimal to (D). Proof. Straightforward, hence omitted. Complementary slackness is studied below using the right-derivative c + j and the left-derivative c j of the function c j defined here. Definition 6. [Right-derivative] (See Section 8A of [22].) The right-derivative of the function c j at a j x j < b j is given by (23) c + j (x j) lim z xj c j (z) c j (x j ) z x j. The right-derivative c + j (b j) is defined to equal. Definition 7. [Left-derivative] (See Section 8A of [22].) The left-derivative of the function c j at a j < x j b j is given by (24) c j (x j) lim z xj c j (z) c j (x j ) z x j. The left-derivative c j (a j) is defined to equal. The reader is referred to Section 8A in [22], Section 24 in [20], and Section 9.1 in [5] for detailed discussions and proofs of the properties of left- and right-derivatives that are summarize next. Since functions c j are convex over [a j, b j ], both the rightand the left-derivatives are finite at every point in the open interval (a j, b j ). The right-derivative can be at the left endpoint a j ; similarly, the left-derivative can be + at the right-endpoint b j. In this context, a flow x X is called regular if c j (x j) < and c + j (x j) > for all j N (see Definition 9.2 in Section 9.1 of [5]). The right-derivative is right-continuous over [a j, b j ) and the left-derivative is left-continuous over (a j, b j ]. Finally, both these derivatives are nondecreasing. Definition 8 (Complementary slackness). (As in Section 9.7 in [5] on finitedimensional monotropic programs.) An x X and a λ Λ are said to be complementary if, for all j N, we have that (25) c j (x j) λ j c + j (x j). 279 278 Remark 9. As in finite-dimensional monotropic programs, an x X and a λ Λ are complementary if and only if x j is the minimizer in (18) for all j N. 280 Proposition 10 (Complementary slackness). Suppose an x F and a λ 281 S are complementary. Then x is optimal to (P ), λ is optimal to (D), and V = W. 282 Suppose x F is optimal to (P ), λ S is optimal to (D), and V = W. Then x 283 and λ are complementary. 284 285 286 287 288 Proof. Suppose x F and λ S are complementary. Since λ S, φ(λ ) = ( ) (26) c j (x j ) λ j x j = c j (x j ) λ j x j = C(x ). }{{} 0 Here, the first equality follows by Remark 9; the second equality holds because the two series converge; finally, the second series is zero because λ S. This implies that

8 A. GHATE 289 x and λ are optimal to (P ) and (D) by Corollary 5 and that their corresponding 290 optimal values are equal. 291 Conversely, suppose that x F and λ S are optimal to (P ) and (D), re- 292 spectively, and that φ(λ ) = C(x ). Since λ S, we have, from Equation (19) that, φ(λ ) = φ j (λ j ) = ( min c j (x j ) λ j x j ). Moreover, C(x ) = c j (x j ) = x j X j c j (x j ) λ j x j = ( ) 294 c j (x j ) λ j x j. Here, the second equality follows because λ j x j = 0 and the third equality holds because the two series converge. Combining these three pieces, we obtain, ( ) c j (x j ) λ j x j = ( ) 296 min c j (x j ) λ j x j. This x j X j ) 297 means that c j (x j ) λ j x j = min c j (x j ) λ j x j for all j N. Remark 9 then 298 299 300 301 302 x j X j ( implies that x and λ are complementary. A planning horizon approach is developed in the next section to establish conditions under which V = W. 4. Absence of a duality gap. For each integer N 1, define the projection of S into R N as 303 304 305 306 307 (27) S N {(x 1, x 2,..., x N ) x S} R N. Note that S N is a (closed) subspace of R N ; by arbitrarily extending S N with a sequence of real numbers, it will, at times, be convenient to view it as a subspace of R N. Viewed this way, S N can be alternatively defined as (28) S N {(x 1, x 2,..., x N ; y N+1, y N+2,...) x S, y R N } R N. 316 308 In order to avoid unnecessary additional notation, S N will denote the above subspace 309 of R N and also the corresponding subspace of R N. Although this is an abuse of 310 notation, the meaning should be clear from context. This practice is standard in 311 countably infinite optimization (see, for example, [27, 28]). 312 Lemma 11. For any integer N 1, S N is a closed subspace of R N. 313 Proof. Let {x(n)} S N R N be a convergent sequence with limit x R N. 314 We show that x S N R N. Since S N is a closed subspace of R N, it is clear 315 that ( x 1, x 2,..., x N ) S N R N. This means that there is some y S such that (y1, y2,..., yn ) = ( x 1, x 2,..., x N ) and ( x 1, x 2,..., x N ; yn+1, y N+2, y N+3,...) 317 S N R N. This shows that ( x 1,..., x N ; x N+1,...) S N R N as per the definition of 318 S N R N in formula (28). 319 Let N n 1, for n = 1, 2,..., be any strictly increasing sequence of positive 320 integers and let S Nn be the corresponding sequence of projections of S. Consider the 321 corresponding sequence of finite-dimensional monotropic optimization problems 322 323 324 325 326 (29) (30) (31) N n (P n ) V n inf C n (x) c j (x j ) x S Nn, x j X j, j = 1, 2,..., N n.

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 9 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 Lemma 12. If (P ) has a feasible solution, then (P n ) has a feasible solution. Proof. If x is feasible to (P ), then its truncation (x 1, x 2,..., x Nn ) R Nn is feasible to (P n ). Let X(n) Nn X j denote the truncation of X to R Nn. Let F n S Nn X(n) R Nn denote the feasible region of (P n ). Lemma 13. If (P n ) has a feasible solution, then it has an optimal solution. Proof. The feasible region F n = S Nn X(n) R Nn of (P n ) equals the intersection of a closed set S Nn R Nn and a compact set X(n) R Nn ; it is therefore compact. The objective in (P n ) is continuous over X(n) and hence also over F n. The result then follows by Corollary 2.35 in [1]. Definition 14. [Kuratowski set limits] Let A 1, A 2,... be a sequence of sets in a metric space Ω. 1. (Lower limit; see Section 29.I on page 335 of [13]) A point p Ω belongs to the lower limit of the sequence {A n } of sets if, for every open ball B(p) with center p, there is some index m large enough such that B(p) A n for all n m. We then write p Li A n. 2. (Upper limit; see Section 29.III on page 337 of [13]) A point p Ω belongs to the upper limit of the sequence {A n } of sets if, for every open ball B(p) with center p, there are infinitely many n with B(p) A n. We then write p Ls A n. 3. (Limit; see Section 29.VI on page 339 of [13]) The sequence {A n } is said to converge to a set A Ω if Li A n = A = Ls A n. We then write Lim A n = A. Remark 15. (See formula 2a in Section 29.VI on page 339 of [13]) With the notation described in the above definitions, if there is a sequence of points p n A n such that p = lim p n, and if the sequence of sets {A n } converges, then p Lim A n. n Remark 16. (See formula 8 in Section 29.VI on page 339 of [13]) With the notation described in the above definitions, if A 1 A 2..., then Lim A n = Ā n, n where Ān denotes the closure of A n. Lemma 17. We have, R N S N1 S N2 S N3.... Moreover, S S Nn. If n S is closed, then S S Nn and hence S = S Nn. n n Proof. For the first claim about set inclusions, fix any integer n 1. Recall as in (28) that S Nn+1 = {(x 1, x 2,..., x Nn+1 ; y Nn+1+1, y Nn+1+2,...) x S, y R N } R N. Thus, if the vector (x 1,..., x Nn+1 ; y Nn+1+1, y Nn+1+2,...) is in S Nn+1, it is also in S Nn. This proves the first claim. For the second claim, suppose x S. Then, by (28), x S Nn R N for each n. Consequently, x S Nn. n For the third claim, suppose x S Nn R N. Thus, for each n, there is a n z(n) S such that (x 1, x 2,..., x Nn ) = (z 1 (n), z 2 (n),..., z Nn (n)) by definition of S Nn. Moreover, the sequence {z(n)} S converges to x. Since S is closed, we have that x S. Proposition 18 (Primal value convergence). Suppose (P ) has a feasible solution and S is closed. Then, problems (P n ) have optimal solutions and lim V n = V. n

10 A. GHATE 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 Proof. The proof is similar to the proof of Berge s maximum principle [4]. Since (P ) has a feasible solution, Lemma 12 implies that (P n ) has a feasible solution. Lemma 13 thus means that (P n ) has an optimal solution, which we denote by x (n). It will, at times, be convenient to view x (n) as belonging to F n = (S Nn X) X R N by seeing S Nn as a subspace of R N. Since the sequence {x (n)} belongs to the compact set X, it has a convergent subsequence. We denote this subsequence by {x (n m )} and use x X to denote its limit. Since x (n m ) S Nnm R N for each m, Remarks 15 and 16; Lemmas 11 and 17; and the assumption that S is closed imply that the limit x S. In other words, x X S = F ; that is, x is feasible to (P ). It is shown next by contradiction that x is optimal to (P ). So suppose not. Since x is not optimal and since (P ) has a feasible solution, there exists a feasible solution ˆx such that c j (ˆx j ) < c j ( x j ). Construct another subsequence { x(n m )} X such that (32) x j (n m ) { x j (n m), j = 1, 2,..., N nm, ˆx j, j = N nm + 1, N nm + 2,.... This new subsequence also converges to x. Thus, the earlier strict inequality can be rewritten as c j (ˆx j ) < c j ( lim x j(n m )). Using continuity of functions c j ( ), m we further rewrite this inequality as c j (ˆx j ) < lim c j( x j (n m )). Now recall m from the proof of Lemma 1 that the series c j (x j ) is uniformly convergent over X. Thus, the series and the limit above can be interchanged (see Theorem 9.7 in [3]) to yield c j (ˆx j ) < lim c j ( x j (n m )). Now splitting each one of the two series m [ N nm Nnm into two parts, we obtain, c j (ˆx j ) + c j (ˆx j ) < lim c j ( x j (n m )) + j=n nm +1 m ] c j ( x j (n m )). Then, substituting for x(n m ) from its definition in (32), we j=n nm +1 [ ] N nm Nnm get, c j (ˆx j ) + c j (ˆx j ) < c j (x j (n m)) + c j (ˆx j ). This j=n nm +1 lim m implies that there exists an m such that (33) That is, N nm N nm c j (ˆx j ) + c j (ˆx j ) < j=n nm +1 N nm c j (ˆx j ) < [ Nnm c j (x j (n m )) + j=n nm +1 j=n nm +1 c j (ˆx j ) c j (x j (n m )). This contradicts the optimality of x (n m ) to (P nm ). This establishes that x is optimal to (P ). To prove value convergence, establish that every convergent subsequence of optimal values V n converges to the same limit and that this limit is V. This would imply that the sequence V n converges to V as required. So let V nm be such a convergent subsequence and let {x (n m )} F nm = (S Nnm X) X R N be the corresponding subsequence of optimal solutions to (P nm ). This subsequence of optimal solutions has a further convergent subsequence, which is denoted by {x (n mt )}. Denote the limit of this convergent subsequence by x. Then, as shown above, x is optimal to (P ). Fix ].

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 11 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 any ɛ > 0. Then H1 implies that there exists a positive integer T large enough such that c j < ɛ/2. Thus, for any t large enough such that N n mt > N nmt, we have, (34) (35) (36) j=n nmt +1 V V nmt = c j ( x j ) < N nmt N nmt N nmt ( c j ( x j ) c j (x j (n mt ))) + 2 c j (x j (n mt )) ( c j ( x j ) c j (x j (n mt ))) + ɛ/2. j=n nmt +1 Now, since lim x t j (n m t ) = x j, and since functions c j ( ) are continuous, there exists a positive integer K large enough such that c j ( x j ) c j (x j (n m t )) < ɛ/(2n nmt ) for all t K. Combining this observation with the inequality in (35), we see that V V nmt < ɛ for all sufficiently large t. Consequently, V nmt converges to V as t. Since V nmt is a further subsequence of the convergent subsequence V nmt, it must converge to the same limit as V nm. In other words, lim = V. Thus, every c j m V n m convergent subsequence of V n converges to V. This completes the proof. (37) (38) The dual of the finite-dimensional monotropic program (P n ) is given by Here, S N n orthogonal to S Nn. N n (D n ) W n sup φ j (λ j ) λ S N n. {λ R Nn Nn } λ j x j = 0, x S Nn is the subspace of R Nn that is Proposition 19 (Zero duality gap, and dual value convergence). Suppose that (P ) has a feasible solution and that S is closed. Suppose W W n for each n. Then V = W. Moreover, lim W n = W. n Proof. Combining weak duality with the hypothesis that W W n yields V W W n. Furthermore, since (P n ) and (D n ) are finite-dimensional monotropic programs, we know that W n = V n (see the zero duality gap result in Proposition 9.19 in [5]). Then using the fact from Proposition 18 that lim V n = V, we get, V W V. n This proves that V = W. For the second part, the aforementioned zero duality gap for finite-dimensional monotropic programs implies that lim W n = lim V n. Then, Propositions 18 and n n the above zero duality gap result imply that lim W n = V = W. n The next lemma shows that W W n if costs are nonnegative. This sufficient condition often holds in network flow problems and in CILPs that arise from infinitehorizon planning applications in Operations Research and Economics (see [8, 9, 10, 11, 17]). Nonnegativity of costs has previously been used to show a zero duality gap

12 A. GHATE 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 result in infinite dimensional linear programs that are more general than CILPs [12]. Lemma 20. Suppose that, for each j N, the cost function c j ( ) is nonnegative over X j. Then W W n, for each n. Proof. For each λ(n) SN n, define ξ(λ(n)) as the corresponding price constructed by appending λ(n) with a sequence of zeros. Specifically, { λ j (n), j = 1, 2,..., N n (39) ξ j (λ(n)) 0, j = N n + 1, N n + 2,.... Observe that ξ(λ(n)) S because for any x S, we have, (40) (41) N n ξ j (λ(n))x j = ξ j (λ(n))x j + N n = λ j (n)x j + +N n ξ j (λ(n))x j +N n ξ j (λ(n))x j = 0. Here, the first sum is zero because λ(n) SN n and the second series is zero because ξ j (λ(n)) = 0 for j = 1 + N n, 2 + N n,.... Now define Ξ n {ξ(λ(n)) λ(n) SN n } S. Consequently, [ N n Nn ) ] (42) W n = sup φ j (λ j (n)) = sup min c j (x j ) λ j (n)x j (43) (44) (45) λ(n) SNn = sup ξ(λ(n)) Ξ n sup ξ(λ(n)) Ξ n sup λ S [ Nn ( min x j X j [ ( min x j X j ( min x j X j [ λ(n) S Nn ( x j X j c j (x j ) ξ j (λ(n))x j ) ] c j (x j ) ξ j (λ(n))x j ) ] c j (x j ) λ j x j ) ] = sup λ S φ j (λ j ) = W. Here, the first equality is simply the definition of W n. The second equality is obtained by substituting for φ j ( ) from (18). The third equality is derived by rewriting the problem equivalently in terms of variables ξ(λ(n)). The first inequality holds because ξ j (λ(n)) = 0 for j = N n + 1, N n + 2,..., and functions c j ( ) are nonnegative. The second inequality holds because the latter problem is a relaxation of the former. The last two equalities in (45) follow by the definition of W. 5. Strong duality. Lemma 21. If (P ) has a regular feasible solution, then finite-dimensional dual problems (D n ) have optimal solutions, denoted by λ (n). Proof. Since (P ) has a regular feasible solution, its projection into R Nn is feasible to (P n ) by Lemma 12 and is also regular. By Lemma 13, (P n ) has an optimal solution, say x (n). Then, by strong duality in finite-dimensional monotropic programs (see Proposition 9.18 from [5]), (D n ) has an optimal solution λ (n) SN n that is complementary to x (n).

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 13 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 Theorem 22 (Strong duality). Suppose that (P ) has a regular feasible solution and that S is closed (thus (P ) has an optimal solution by Lemma 1). Let {x (n), λ (n)}, for n = 1, 2,..., be a sequence of pairs of complementary optimal solutions to (P n ) and (D n ) as in Lemmas 13 and 21. Suppose there exists a sequence v Λ such that λ j (n) v j for all j N and all n. Suppose S is closed in the product topology on Λ. Then (D) has an optimal solution (with W = V ). Proof. Define the set Z [ v j, v j ], which is in Λ by H2 and H3 because v Λ. This Z is compact by Tychonoff s theorem. By appending solutions λ (n) with a sequence of zeros, we view them as elements of subspace Λ, and in particular, as members of the set Z. As in the proof of Proposition 18, we continue to use x (n) to also denote its extension in F n = (S Nn X) X R N. Since X and Z are compact, the sequence {x (n), λ (n)} X Z has a convergent subsequence {x (n k ), λ (n k )}. We denote its limit by ( x, λ) X Z. As shown in Proposition 18, x F. Since Z Λ, we know that λ Λ. More specifically, we prove that λ S. To see this, note that the sequence {λ (n)} Λ in fact belongs to S. This holds because, for any y S, λ j (n)y j = Nn λ j (n)y j + λ j (n)y j = 0. Here, the last equality holds because N n j=n n+1 λ j (n)y j = 0 as (λ 1(n), λ 2(n),..., λ N n (n)) SN n, and λ j (n) = 0 for j N n + 1. Since S is assumed to be closed, the limit λ of the sequence {λ (n)} S must belong to S. It is shown next that x F and λ S are complementary in (P ) and (D). This implies, by the first claim in Proposition 10, that λ is optimal to (D). To show that x and λ are complementary, we need to prove that (46) c j ( x j) λ j c + j ( x j), for every j N. If j is such that the interval X j is a single point, that is, if a j = b j, then x j = a j = b j. Consequently, c j ( x j) = and c + j ( x j) = ; the inequality in (46) then holds for any real number λ j. We therefore only need to focus on indices j for which the interval X j is not a single point. For these indices, we prove (46) by contradiction. So that there exists either a j N such that (47) c j ( x j) > λ j or a j N such that (48) c + j ( x j) < λ j. 507 503 First suppose that it is the former. Then x j a j because c j (a j) = and 504 this contradicts (47). We consider two other subcases: (A) there exists a subsequence 505 x j (n k t ) such that x j (n k t ) x j for all t; and (B) there is no such subsequence and 506 hence x j (n k) > x j for all k large enough. In subcase (A), x j (n k t ) x j as t. Then, since the left-derivative is left-continuous over (a j, b j ], we have, c j ( x j) = c j ( 508 x j (n k t )) = lim c t j (x j (n k t )). Moreover, λ j = lim λ t j (n k t ). These two observations, 509 when combined with (47), yield that lim c t j (x j (n k t )) > lim t λ j (n k t ). Thus there 510 must be some t for which c j (x j (n k t )) > λ j (n k t ). But this contradicts the fact that

14 A. GHATE 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 the pair x j (n k t ), λ j (n k t ) is complementary in (P nkt ) and (D nkt ). In subcase (B), we have that c j (x j (n k)) c j ( x j) for all k large enough because the left-derivative is nondecreasing. Moreover, lim k λ j (n k) = λ j. Combining these two observations with (47) we obtain c j (x j (n k)) > λ j (n k) for some k large enough. Again, this contradicts the fact that the pair x j (n k), λ j (n k) is complementary in (P nk ) and (D nk ). Now suppose that it is the latter. Then x j b j because c + j (b j) = + and this contradicts (48). We consider two other subcases: (A) there exists a subsequence x j (n k t ) such that x j (n k t ) x j for all t; and (B) there is no such subsequence and hence x j (n k) < x j for all k large enough. In subcase (A), x j (n k t ) x j as t. Then, since the right-derivative is right-continuous over [a j, b j ), we have, c + j ( x j) = c + j ( x j (n k t )) = lim c + t j (x j (n k t )). Moreover, λ j = lim λ t j (n k t ). These two observations, when combined with (48), yield that lim c + t j (x j (n k t )) < lim λ t j (n k t ). Thus there must be some t for which c + j (x j (n k t )) < λ j (n k t ). But this contradicts the fact that the pair x j (n k t ), λ j (n k t ) is complementary in (P nkt ) and (D nkt ). In subcase (B), we have that c + j (x j (n k)) c + j ( x j) for all k large enough because the right-derivative is nondecreasing. Moreover, lim k λ j (n k) = λ j. Combining these two observations with (48) we obtain c + j (x j (n k)) < λ j (n k) for some k large enough. Again, this contradicts the fact that the pair x j (n k), λ j (n k) is complementary in (P nk ) and (D nk ). 6. Countably infinite linear programs. As in [25], consider the following class of linear optimization problems: 532 (49) (CILP) inf c j x j 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 (50) d ij x j = b i, i = 1, 2,..., (51) 0 x j u j, j = 1, 2,.... These problems can be seen as special cases of the CILPs considered in [9, 24] wherein the formulation did not include explicit upper bounds on variables but nevertheless required the variables to be bounded without loss of optimality for strong duality. We work with the bounded-variable version here to be consistent with our framework in Section 2. Here, c R N, b R N, and 0 < u R N are given sequences of real numbers and D = {d ij } is a doubly-infinite matrix of real numbers. As in [24, 25], assume that each row and each column of D includes a finite number of nonzero entries. This assumption is prevalent in duality results on CILPs and it holds in most problems in Operations Research. Under this structure of the constraint coefficient matrix, assume without loss of generality that there is an increasing sequence of positive integers {M n } with the following property: variables x j indexed by j = M n + 1, M n + 2,... do not appear in the equality constraints (50) indexed by i = 1, 2,..., n. Thus, the ith equality constraints (50) can be equivalently written as Mi d ij x j = b i. We first convert this problem into a countably infinite monotropic program by employing a process identical to that in finite-dimensional LPs (see Section 9.7 in [5]). Introduce the sequence z = (z 1, z 2,...) of variables to rewrite the ith constraint in (50) as Mi d ij x j z i = 0, and add constraints b i z i b i, for i = 1, 2,.... Use the

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 15 553 554 555 556 shorthand (x, z) to denote sequences (x 1, x 2... ; z 1, z 2,...), and consider the subspace { (52) S (x, z) M i } d ij x j z i = 0, i = 1, 2,.... Also let intervals X j [0, u j ] for j N and Z i [b i, b i ]. Then, problem (CILP) can be equivalently rewritten as the countably infinite monotropic program 557 (53) (MONO-CILP) inf c j x j 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 (54) (55) (56) (x, z) S, x j X j, j = 1, 2,..., z i Z i, i = 1, 2,.... Hypothesis H1 from Section 2 then reduces to: H1. the series c j x j of nonnegative terms is finite. max x j X j Several applications where H1 holds, including CILP formulations of infinite-horizon planning problems, countable state MDPs, and robust optimization, are discussed in [17, 10, 11, 24, 25]. The space Λ of dual sequences λ = (λ 1, λ 2,...) and the space Θ of dual sequences θ = (θ 1, θ 2,...) are then charaterized by the hypotheses H2. the series λ j u j + θ i b i of nonnegative terms is finite; and i N H3. the series λ j y j + θ i w i of nonnegative terms is finite for every (y, w) i N S. Let X X j and Z Z i. The Lagrangian function in (14) then reduces to i N (57) L((x, z), (y, w); λ, θ) = (c j x j λ j x j ) i N θ i z i + λ j y j + i N θ i w i, 574 575 576 577 578 for x X, z Z, (y, w) S, λ Λ, and θ Θ. Now, similar to (12), we define, (58) (59) φ j (λ j ) min x j X j (c j x j λ j x j ), ψ i (θ i ) min z i Z i θ i z i = θ i b i, because Z i = {b i }. Then the dual problem (D) from Section 2 reduces to 579 580 581 (60) (61) (D-MONO-CILP) sup φ j (λ j ) b i θ i i N (λ, θ) S, 582 583 584 585 where { (62) S (λ, θ) Λ Θ λ j y j + i N } θ i w i = 0, (y, w) S. The weak duality result in Proposition 4 and complementary slackness result in Proposition 10 then hold for the pair (MONO-CILP) and (D-MONO-CILP).

16 A. GHATE 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 In order to apply our zero duality gap and strong duality results from Sections 4 and 5, we first define appropriate finite-dimensional projections of (MONO-CILP). Toward this end, let N n M n +n, for all positive integers n. We define the projection subspaces } (x, (63) S Nn {(x 1,..., x Mn ; z 1,..., z n ) R Nn z) S. Lemma 23. Consider the subspace M n } (64) σ Nn {(x 1,..., x Mn ; z 1,..., z n ) R Nn d ij x j z i = 0, i = 1, 2,..., n. Then, S Nn = σ Nn. Proof. First suppose that (x 1, x 2,..., x Mn ; z 1, z 2,..., z n ) S Nn. Then, (x, z) S and hence Mn d ij x j z i = 0 for all i; in particular, Mn d ij x j z i = 0 for i = 1, 2,..., n. Thus, the vector (x 1, x 2,..., x Mn ; z 1, z 2,..., z n ) σ Nn. This shows that S Nn σ Nn. Now suppose that (x 1, x 2,..., x Mn ; z 1, z 2,..., z n ) σ Nn. We use this to construct a solution (x 1,..., x Mn, x Mn+1, x Mn+2,... ; z 1,..., z n, z n+1, z n+2,...) S as follows. Set x Mn+1 = x Mn+2 =... = 0 and z n+k = Mn d (n+k)j x j. As a result, the vector (x 1, x 2,..., x Mn ; z 1, z 2,..., z n ) S Nn. Thus, σ Nn S Nn. These two observations show that S Nn = σ Nn. In view of the above lemma, we equivalently use σ Nn instead of S Nn to formulate the finite-dimensional monotropic programming projections of (MONO-CILP) as 605 606 607 608 609 (65) (66) (67) (68) M n (MONO-CILP(n)) inf c j x j M n d ij x j z i = 0, i = 1, 2,..., n, 0 x j u j, j = 1, 2,..., M n, z i = b i, i = 1, 2,..., n. 610 611 612 613 614 As a side note, these projections are familiar, finite-dimensional LPs; this is the benefit of employing σ Nn instead of S Nn to write these problems. We have, Lemma 24. The subspace S defined in (52) is closed. Proof. Consider a convergent sequence {(x n, z n )} S with limit ( x, z). We need to show that ( x, z) S. Consider any fixed i. We have, 615 616 (69) M i M i ( ) d ij x j z i = d ij lim n xn j lim n zn i ( M i = lim n ) d ij x n j zi n = 0. 617 618 Here, the last equality holds because (x n, z n ) S. Since i was arbitrary, ( x, z) S as required.

DUALITY IN COUNTABLY INFINITE MONOTROPIC PROGRAMS 17 619 620 621 622 623 Thus, as long as (MONO-CILP) has a feasible solution, optimal values of the finitedimensional projections (MONO-CILP(n)) converge to the optimal value of (MONO- CILP) as per Proposition 18. We also note, from the fundamental theorem of linear n algebra (see page 198 of [30]), that σn n = {(λ 1,..., λ Mn ; θ 1,..., θ n ) d ij θ i + λ j = 0, j = 1,..., M n }. Thus, the dual of (MONO-CILP(n)) is given by i=1 624 625 626 (70) (71) M n n (D-MONO-CILP(n)) sup φ j (λ j ) b i θ i i=1 n d ij θ i + λ j = 0, j = 1,..., M n. i=1 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 If c j 0 for all j N, then optimal values of these finite-dimensional duals provide a lower bound on the optimal value of (D-MONO-CILP) as per Lemma 20. Thus, in this case, if (MONO-CILP) has a feasible solution, then there is no duality gap between (MONO-CILP) and (D-MONO-CILP) by Proposition 19. The remaining two hypotheses from the strong duality result in Theorem 22 require that the optimal variables in (D-MONO-CILP(n)) be bounded independently of n and the orthogonal subspace S be closed. We therefore first characterize S and show that it is closed. We need additional notation to achieve this. So recall that each column of D includes a finite number of nonzero entries. For each column j, we thus use I j to denote the finite set of rows i for which a ij 0. Also recall that each row of D includes a finite number of nonzero entries. Thus, for each row i, let J i denote the finite set of columns j for which a ij 0. Starting with J 0 =, we define J i 1 i 1 J i, for i = 1, 2,.... Thus, Ji, for i = 0, 1, 2,... is the set of k=1 variables that appear in the first i equality constraints. Consequently, K i J i \ J i 1, for i = 1, 2,..., is the set of variables that appear in the first i equality constraints but do not appear in the first i 1 equality constraints. Lemma 25. The orthogonal subspace S in (62) can be equivalently rewritten as { (72) S } = (λ, θ) Λ Θ d ij θ i + λ j = 0, j = 1, 2,.... i I j Proof. Suppose (λ, θ) S as defined in (72). Let (y, w) S. Then, we have, (73) λ j y j + θ i w i = λ j y j + θ i w i i N i N j J i\ J i 1 i N = (74) i N j J i\ J i 1 = ( (75) θ i d ij y j i N j J i ( k Ij d kj θ k ) y j + i N θ i w i ) + θ i w i = θ i w i + θ i w i = 0. i N i N i N Here, the first equality is obtained by reordering terms in the series λ j y j, which is allowed because the series converges absolutely by H3. The second equality holds

652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 18 A. GHATE because λ j = d kj θ k as per (72). The third equality is obtained by collecting k I j terms according to their indices i. This shows that (λ, θ) S as defined in (62). Now, for the reverse, suppose that (λ, θ) S as defined in (62). We will show by contradiction that (λ, θ) S as defined in (72). So suppose not. Then, there must exist some j, say j = 1 without loss of generality, such that λ 1 + d i1 θ i = ɛ 1 0. i I 1 Now consider a (y, w) defined by y 1 = 1, y j = 0 for j 1; and w i = d i1 for all i. This implies that (y, w) S because, for any i, we have, d ij y j = d i1 = w i. j J i Moreover, we obtain, λ j y j + θ i w i = λ 1 + θ i d i1 = λ 1 + θ i d i1 = ɛ 1 0. i N i N i I 1 This contradicts the fact that (λ, θ) S as defined in (62). This lemma can be viewed as a countably infinite extension of one piece of the fundamental theorem of linear algebra and hence it may be of independent interest. This lemma allows us to view the constraint (λ, θ) S in (D-MONO-CILP) in the more famililar manner as d ij θ i + λ j = 0, for j = 1, 2,..., using the transpose of i I j the doubly infinite matrix D. Lemma 26. The orthogonal subspace defined in (72) is closed. Proof. Suppose {(λ n, θ n )} S is a convergent sequence with limit ( λ, θ). We need to show that ( λ, θ) S. Fix any j. We have, (76) d ij θi + λ j = i I j i I j d ij ( ) ( ) lim n θn i + lim n λn j = lim d ij θi n + λ n j = 0. n i I j Here, the first equality holds by definition of θ and λ. The second equality holds because the set I j is finite. The last equality holds because (λ n, θ n ) S. Since j was arbitrary, ( λ, θ) S. Thus, Theorem 22 implies that as long as (MONO-CILP) has a feasible solution, strong duality holds between (MONO-CILP) and (D-MONO-CILP) when optimal dual variables in (D-MONO-CILP(n)) are bounded independently of n. This essentially recovers Theorem 3.2 in [9]. See [9, 24, 25] for several applications where variables in the finite-dimensional dual problems are bounded in this way without loss of optimality. 7. Convex minimum cost flow problems on infinite networks. Recall the description of a minimum cost flow problem in countably infinite networks from the author s recent work in [17]. Let N denote a countable set of nodes. An arc from node i N to node j N is denoted by the ordered tuple (i, j); let A denote the countable set of all such arcs. Assume the standard regularity condition that each node has finite in- and out-degree (see [23, 29]). The resulting infinite directed network is denoted by G (N, A). A real number s i, called source, is assigned to each node i N. If s i > 0, node i is called a supply node and a net flow of s i needs to be pushed out from this node; if s i = 0, node i is called a transshipment node; finally, if s i < 0, node i is called a demand node and a net flow of s i needs to be delivered to this node. The largest amount of flow that can be carried through arc (i, j) A is denoted by 0 < u ij < ; these are called flow capacities. Arc flows are denoted by x ij, for (i, j) A. For each arc (i, j) A, let c ij : [0, u ij ] R be a real-valued, continuous, and convex function.