Abstract
We study the polyhedral structure of a mixed 0-1 set arising from the submodular maximization problem, given by \(P = \{(w,x)\in {\mathbb {R}}\times \{0,1\}^n: w\le f(x), x\in {\mathcal {X}}\}\), where submodular function f(x) is represented by a concave function composed with an affine function, and \({\mathcal {X}}\) is the feasible region of binary variables x. For \({\mathcal {X}}= \{0,1\}^n\), two families of facet-defining inequalities are proposed for the convex hull of P through restriction and lifting using submodular inequalities. When \({\mathcal {X}}\) involves multiple disjoint cardinality constraints, we propose a new class of facet-defining inequalities for the convex hull of P through multidimensional sequence independent lifting. The derived polyhedral results not only strengthen and generalize some existing developments in the literature, but are also linked to the classical results for the mixed 0-1 knapsack and single-node flow sets. Our computational study on a set of randomly generated submodular maximization instances demonstrates the superiority of the proposed facet-defining inequalities within a branch-and-cut scheme.
Similar content being viewed by others
References
Ahmed, S., Atamtürk, A.: Maximizing a class of submodular utility functions. Math. Program. 128(1–2), 149–169 (2011)
Angulo, A., Espinoza, D., Palma, R.: Sequence independent lifting for mixed knapsack problems with gub constraints. Math. Program. 154(1–2), 55–80 (2015)
Atamtürk, A.: Sequence independent lifting for mixed-integer programming. Oper. Res. 52(3), 487–490 (2004)
Atamtürk, A., Gómez, A.: Maximizing a class of utility functions over the vertices of a polytope. Oper. Res. 65(2), 433–445 (2017)
Atamtürk, A., Küçükyavuz, S., Tezel, B.: Path cover and path pack inequalities for the capacitated fixed-charge network flow problem. SIAM J. Optim. 27(3), 1943–1976 (2017)
Atamtürk, A., Narayanan, V.: Polymatroids and mean-risk minimization in discrete optimization. Oper. Res. Lett. 36(5), 618–622 (2008)
Atamtürk, A., Narayanan, V.: The submodular knapsack polytope. Discret. Optim. 6(4), 333–344 (2009)
Buchbinder, N., Feldman, M.: Submodular functions maximization problems. In: Handbook of Approximation Algorithms and Metaheuristics, Second Edition, pp. 771–806. CRC Press (2018)
Chen, W., Wang, Y., Yuan, Y.: Combinatorial multi-armed bandit: General framework and applications. In: International Conference on Machine Learning, pp. 151–159 (2013)
Conforti, M., Cornuéjols, G., Zambelli, G.: Integer Programming. Springer, Berlin (2014)
Dolhansky, B.W., Bilmes, J.A.: Deep submodular functions: Definitions and learning. In: Advances in Neural Information Processing Systems, pp. 3404–3412 (2016)
El-Arini, K., Veda, G., Shahaf, D., Guestrin, C.: Turning down the noise in the blogosphere. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 289–298. ACM (2009)
Fujishige, S.: Submodular Functions and Optimization. Elsevier, Netherlands (2005)
Gu, Z., Nemhauser, G.L., Savelsbergh, M.W.: Sequence independent lifting in mixed integer programming. J. Comb. Optim. 4(1), 109–129 (2000)
Gurobi Optimization, L.: Gurobi optimizer reference manual (2019). http://www.gurobi.com
Itoko, T., Iwata, S.: Computational geometric approach to submodular function minimization for multiclass queueing systems. In: International Conference on Integer Programming and Combinatorial Optimization, pp. 267–279. Springer (2007)
Kılınç-Karzan, F., Küçükyavuz, S., Lee, D.: Joint chance-constrained programs and the intersection of mixing sets through a submodularity lens. Mathematical Programming pp. 1–44 (2021)
Krause, A., Golovin, D.: Submodular function maximization. In: Bordeaux, L., Hamadi, Y., Kohli, P. (eds.) Tractability: Practical Approaches to Hard Problems, pp. 71–104. Cambridge University Press, Cambridge (2014)
Li, J., Deshpande, A.: Maximizing expected utility for stochastic combinatorial optimization problems. Math. Oper. Res. 44(1), 354–375 (2019)
Lovász, L.: Submodular functions and convexity. In: Mathematical Programming The State of the Art, pp. 235–257. Springer (1983)
Marchand, H., Wolsey, L.A.: The 0–1 knapsack problem with a single continuous variable. Math. Program. 85(1), 15–33 (1999)
Nemhauser, G.L., Wolsey, L.A.: Integer and Combinatorial Optimization. Wiley, New York (1988)
Nemhauser, G.L., Wolsey, L.A., Fisher, M.L.: An analysis of approximations for maximizing submodular set functions-I. Math. Program. 14(1), 265–294 (1978)
Rusmevichientong, P., Shen, Z.J.M., Shmoys, D.B.: Dynamic assortment optimization with a multinomial logit choice model and capacity constraint. Oper. Res. 58(6), 1666–1680 (2010)
Schoemaker, P.J.: The expected utility model: Its variants, purposes, evidence and limitations. J. Econ. Literat. pp. 529–563 (1982)
Shi, X., Prokopyev, O., Zeng, B.: Sequence independent lifting for the set of submodular maximization problem. In: International Conference on Integer Programming and Combinatorial Optimization. Springer (2020)
Stobbe, P., Krause, A.: Efficient minimization of decomposable submodular functions. In: Advances in Neural Information Processing Systems, pp. 2208–2216 (2010)
Tschiatschek, S., Iyer, R.K., Wei, H., Bilmes, J.A.: Learning mixtures of submodular functions for image collection summarization. In: Advances in Neural Information Processing Systems, pp. 1413–1421 (2014)
Wei, K., Iyer, R., Bilmes, J.: Submodularity in data subset selection and active learning. In: International Conference on Machine Learning, pp. 1954–1963 (2015)
Wolsey, L.A.: Valid inequalities and superadditivity for 0–1 integer programs. Math. Oper. Res. 2(1), 66–77 (1977)
Wolsey, L.A.: Submodularity and valid inequalities in capacitated fixed charge networks. Oper. Res. Lett. 8(3), 119–124 (1989)
Wu, H.H., Küçükyavuz, S.: A two-stage stochastic programming approach for influence maximization in social networks. Comput. Optim. Appl. 69(3), 563–595 (2018)
Wu, H.H., Küçükyavuz, S.: An exact method for constrained maximization of the conditional value-at-risk of a class of stochastic submodular functions. Oper. Res. Lett. 48(3), 356–361 (2020)
Yu, J., Ahmed, S.: Maximizing expected utility over a knapsack constraint. Oper. Res. Lett. 44(2), 180–185 (2016)
Yu, J., Ahmed, S.: Maximizing a class of submodular utility functions with constraints. Math. Program. 162(1–2), 145–164 (2017)
Yu, J., Ahmed, S.: Polyhedral results for a class of cardinality constrained submodular minimization problems. Discret. Optim. 24, 87–102 (2017)
Yue, Y., Guestrin, C.: Linear submodular bandits and their application to diversified retrieval. In: Advances in Neural Information Processing Systems, pp. 2483–2491 (2011)
Zeng, B., Richard, J.P.P.: A framework to derive multidimensional superadditive lifting functions and its applications. In: Fischetti, M., Williamson, D.P. (eds.) Integer Programming and Combinatorial Optimization, pp. 210–224. Springer, Heidelberg (2007)
Zeng, B., Richard, J.P.P.: A polyhedral study on 0–1 knapsack problems with disjoint cardinality constraints: strong valid inequalities by sequence-independent lifting. Discret. Optim. 8(2), 259–276 (2011)
Acknowledgements
The authors are thankful to the Review Team for their detailed and constructive comments.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
An extended abstract of this work appeared in the proceedings of the 21st Conference on Integer Programming and Combinatorial Optimization (IPCO 2020) [26]
Appendices
Appendix
A Proof of Theorem 6
Let us first consider the function \(\gamma _T(z)\) (defined in equation (22)). As shown in Example 3, \(\gamma _T(z)\) is not subadditive on \(z\in {\mathbb {R}}_+\) in general for \(T \ne \emptyset \). However, we can still find the subadditive structure inside \(\gamma _T(z)\) in the following discussion. Consider a new function \({\tilde{\gamma }}_T(z)\) as follows:
Recall that we assume \(S = \{1, \ldots , s\}\) such that \(a_1\ge \cdots \ge a_s\). Let \({\tilde{A}}_k = \sum _{j\in [k]\setminus T} a_j\) for \(k\in S\) and \({\tilde{A}}_0 = 0\). Note that if \(k \in T\), then \({\tilde{A}}_k = {\tilde{A}}_{k-1}\). Then following Lemma 8, we have \({\tilde{\gamma }}_T({\tilde{A}}_k) = \sum _{j\in [k]\setminus T} \rho _j(S\setminus j)\) and
Consequently, based on Proposition 4, we can verify that function \({\tilde{\gamma }}_T(z)\) is subadditive on \(z \in {\mathbb {R}}_+\) for a given \(T\subseteq S\).
To establish the proof of Theorem 6, our basic idea is to exploit inequalities (20) in Lemma 7. To show (20), we first establish that inequalities similar in spirit to (20) hold for \(\gamma _T(z)\) in Lemmata 13 and 14, which, in turn, rely on the technical results in Lemmata 11 and 12. Then we use the fact that \(\gamma {\left( {\begin{array}{c}z\\ {\mathbf {u}}\end{array}}\right) }\) can be computed through equation (23) and \(\gamma _T(z)\) to complete the proof of Theorem 6.
Lemma 11
Function \({\tilde{\gamma }}_T(z)\) is nonincreasing on T, that is \({\tilde{\gamma }}_T(z) \le {\tilde{\gamma }}_{T'}(z) \le {\gamma _0(z)} \) for any \(T'\subseteq T \subseteq S.\)
Proof
Note that if \(T=\emptyset \), then \({\tilde{\gamma }}_\emptyset (z) = \gamma _\emptyset (z) = \gamma _0(z)\). Thus, it suffices to show that \({\tilde{\gamma }}_T(z) \le {\tilde{\gamma }}_{T\setminus j_0}(z)\) for any \(j_0\in T\). Let \(T' = T\setminus j_0\) and \({\tilde{A}}'_k = \sum _{j\in [k]\setminus T'} a_j\) for \(k\in S\).
If \(z \le {\tilde{A}}_{j_0-1}\), then \({\tilde{A}}'_k = {\tilde{A}}_k\) for \(k=0, \ldots , j_0-1\). By (29), it is easy to verify that \({\tilde{\gamma }}_T(z) = {\tilde{\gamma }}_{T'}(z)\).
If \(z > {\tilde{A}}_{j_0-1}\), then \({\tilde{A}}'_k = {\tilde{A}}_k + a_{j_0}\) for \(k=j_0, \ldots , s\). By (29), it is easy to verify that \( {\tilde{\gamma }}_{T'}(z+a_{j_0}) - {\tilde{\gamma }}_{T'}({\tilde{A}}'_{j_0}) = {\tilde{\gamma }}_T(z) - {\tilde{\gamma }}_T({\tilde{A}}_{j_0-1}). \) Meantime, we can verify that \({\tilde{\gamma }}_{T'}(z)\) has the form of (12) in Sect. 2.3. Then by Lemma 2 we have
which implies that \({\tilde{\gamma }}_{T'}(z) \ge {\tilde{\gamma }}_T(z)\) as \({\tilde{\gamma }}_{T'}({\tilde{A}}'_{j_0-1}) = {\tilde{\gamma }}_T({\tilde{A}}_{j_0-1})\). \(\square \)
Lemma 12
Let \(\varDelta \in [0, a_j]\) for some \(j\in T\). If \(0\le z \le {\tilde{A}}_{j-1}\), then
Proof
Firstly, by Lemma 8, we have
We assume that \(z \in [{\tilde{A}}_{k-1}, {\tilde{A}}_{k}]\) for some \(k\in S\setminus T\), then \(k \le j\) and \(a_k \ge a_j\). If \(z - \varDelta \le 0\), then \(k=1\). Let \(\varOmega = a(S) + z \ge a(S)\). By Lemma 8, we have
where the inequality follows from \(a_1 \ge a_j\ge \varDelta \) and the concavity of g.
If \(z - \varDelta \ge 0\), then \( \gamma _T(z+a(T)) - \gamma _T(z+a(T)-\varDelta ) ={\tilde{\gamma }}_T(z) - {\tilde{\gamma }}_T(z-\varDelta )\) based on the definition of \({\tilde{\gamma }}_T(z)\). Assume \(z - \varDelta \in [{\tilde{A}}_{\ell -1}, {\tilde{A}}_{\ell }]\) for some \(\ell \le k\). Since \(\varDelta \le a_j \le a_{k}\), then either \(\ell = k\) or \(\ell =k-1\). If \(\ell = k\), then
where the inequality follows from \(z \le {\tilde{A}}_{k}\) and the concavity of g. If \(\ell =k-1\), let \(\varOmega = a(S) - {\tilde{A}}_{k-1} + z \ge a(S)\). By equation (29), we have
where the inequality follows from \(\varDelta \le a_k\) and the concavity of g. \(\square \)
Lemma 13
For any \(T\subseteq S\), we have
where \(z(T) = \sum _{j\in T} z_j.\)
Proof
We prove the result by induction. If \(|T| = 1\), then the statement is trivial.
If the statement holds for \(|T| = k-1\), then we establish that the statement still holds for \(|T| = k\). Observe that it is sufficient to show that there exists some \(\ell \in T\) such that
where \(T' = T\setminus \ell \). There are four possible cases to consider:
Case 1 \(\exists \; \ell \in T\) such that \(z_\ell \ge a_\ell , z(T') \ge a(T')\):
Based on the assumption, we have \(z(T) \ge a(T)\), it follows that
where the first inequality is based on Lemma 11, and the second inequality follows from the fact that \({\tilde{\gamma }}_T(z)\) is subadditive on \(z\in {\mathbb {R}}_+\) for a given T.
Case 2 \(\exists \; \ell \in T\) such that \(z_\ell \le a_\ell , z(T') \le a(T')\):
Based on the assumption, we have \(z(T) \le a(T)\). Let \(\varOmega = a(S) + z(T') - a(T')\), then \(\varOmega \le a(S)\). By Lemma 8, it follows that
where the inequality follows from the concavity of g and \(z_\ell - a_\ell \le 0.\)
Case 3 \(\exists \; \ell \in T\) such that \(z_\ell \le a_\ell , z(T') \ge a(T')\):
Based on Lemma 8, we have \(\gamma _{\{\ell \}}(z_\ell ) = g(a(S) - a_\ell + z_\ell ) - g(a(S) - a_\ell )\). Without loss of generality, we assume \(\ell \) is the smallest index in T. Then there have two cases to consider:
-
if \(z(T') \ge a(T') + {\tilde{A}}_{\ell -1}\), we have
$$\begin{aligned}&\gamma _{T'}(z_\ell + z(T')) - \gamma _{T'}(z(T')) \\&\quad = {\tilde{\gamma }}_{T'}(z(T')-a(T')+z_\ell ) - {\tilde{\gamma }}_{T'}(z(T')-a(T')) \\&\quad \le {\tilde{\gamma }}_{T'}(A_{\ell -1} + z_\ell ) - {\tilde{\gamma }}_{T'}( A_{\ell -1}) = \gamma _{\{\ell \}}(z_\ell ), \end{aligned}$$where the inequality follows from Lemma 2 that \({\tilde{\gamma }}\) has the form of (11). By Lemma 11, we have that
$$\begin{aligned} \gamma _{\{\ell \}}(z_\ell ) + \gamma _{T'}(z(T')) \ge \gamma _{T'}(z_\ell +z(T')) \ge \gamma _T(z(T)). \end{aligned}$$ -
if \(z(T') \le a(T') + {\tilde{A}}_{\ell -1}\), then by Lemma 8 it can be verified that
$$\begin{aligned} \gamma _T(z(T') + a_\ell ) = \gamma _{T'}(z(T')) + \rho _\ell (S\setminus \ell ). \end{aligned}$$Let \(\varDelta = a_\ell - z_\ell \ge 0\). Note that \(\varDelta \le a_\ell \) and \(z(T')-a(T')\le {\tilde{A}}_{\ell -1}\), it follows from Lemma 12 that
$$\begin{aligned} \gamma _T(z(T') + a_\ell ) - \gamma _T(z(T')+z_\ell )&\ge \gamma _T(a(T)) - \gamma _T(a(T)-\varDelta ) \\&= g(a(S)) - g(a(S)-\varDelta ) \\&= \rho _\ell (S\setminus \ell ) - \gamma _{\{\ell \}}(z_\ell ). \end{aligned}$$Replacing \(\gamma _T(z(T') + a_\ell )\) with \(\gamma _{T'}(z(T')) + \rho _\ell (S\setminus \ell )\) in the above inequality, it yields the desired results.
Case 4 \(\exists \; \ell \in T\) such that \(z_\ell \ge a_\ell , z(T') \le a(T')\):
Since \(z(T') \le a(T')\), then there exists \(\ell ' \in T\) such that \(z_{\ell '} \le a_{\ell '}\). We can reduce this case to either Case 2 or Case 3.
In summary, the claim holds for \(|T| = k\) and we complete the proof. \(\square \)
Lemma 14
For any \(j\in S\) and \(z_0, z_1 \ge 0\), we have
Proof
If \(z_1 \ge A_{j-1}\), then \(\gamma _{\{j\}}(z) = \gamma _0(z)\) and \(\gamma _{\{j\}}(z_0+z_1) = \gamma _0(z_0+z_1)\) based on Lemma 8. It yields that the claim holds due to the subadditivity of \(\gamma _0\) in this case.
If \(a_{j} \le z_1 \le A_{j-1}\), then \(\gamma _{\{j\}}(z_1) = {\tilde{\gamma }}(z_1 - a_{j}) + \rho _{j}(S\setminus j)\). Note that \(\gamma _0(z) = {\tilde{\gamma }}_{\{j\}}(z)\) when \(z \le A_{j-1}\) through (29), thus
where the second inequality follows from Lemma 11 that \({\tilde{\gamma }}_{\{j\}}(z) \le \gamma _0(z)\).
If \(z_1 \le a_{j}\) and \(z_0+z_1 \le a_{j}\), then by Lemma 8, we have
where the first inequality follows from \(a_{j} \le a_1\), and the second inequality follows from the subadditivity of \(\gamma _0\).
If \(z_1 \le a_{j}\), \(z_0+z_1 \ge a_j\) and \( z_0 \le A_{j}\), then let \(\varDelta = a_{j} - z_1 \le z_0\). By Lemmata 1 and 8, we have that
Let \(z_1'=a_{j}\) and \(z_0' = z_0-\varDelta \), then \(z_0'+z_1' = z_0 + z_1\). Since \(a_{j}\le z_1' \le A_{j}\), based on the aforementioned case, we have
Summing the above two inequalities, we get the desired result.
If \(z_1 \le a_{j}\) and \(z_0\ge A_{j-1}\), then by Lemma 8, we have \(\gamma _{\{j\}}(z_1) = \gamma _0(A_{j-1} + z_1) - \gamma _0(A_{j-1})\) and \(\gamma _{\{j\}}(z_0+z_1) = \gamma _0(z_0+z_1)\). Then the result directly follows from Lemma 2. \(\square \)
We now come to prove Theorem 6.
Proof
(Theorem 6) It suffices to show that the lifting function \(\gamma {\left( {\begin{array}{c}z\\ {\mathbf {u}}\end{array}}\right) }\) satisfies the SI condition (20). Let \(\varGamma \subseteq {\bar{S}}\) and \(z_j \ge 0\) for \(j\in \varGamma \). Then for any \(T\subseteq S\) such that \(|T_i| = \max \{0, |S_i|+ |\varGamma _i| -d_i\}\) for all \(i\in [r]\), we construct \(\{T^j\}_{j\in \varGamma }\) as follows: for each \(i\in [r]\), suppose \(\varGamma _i = \{1, \ldots , |\varGamma _i|\}\) and \(T_i = \{i_1, \ldots , i_{|T_i|}\}\),
-
if \(|S_i| < d_i\), then \(|T_i| < |\varGamma _i|\). Let \(T^j = \{i_j\}\) for \(j=1,\ldots , |T_i|\), and \(T^j = \emptyset \) for \(j=|T_i|+1, \ldots , |\varGamma _i|\).
-
if \(|S_i| = d_i\), then \(|T_i| = |\varGamma _i|\). Let \(T^j = \{i_j\}\) for \(j\in \varGamma _i\).
Therefore, by Lemmata 13 and 14, we have
Recall equation (23), we have
Rights and permissions
About this article
Cite this article
Shi, X., Prokopyev, O.A. & Zeng, B. Sequence independent lifting for a set of submodular maximization problems. Math. Program. 196, 69–114 (2022). https://doi.org/10.1007/s10107-022-01801-y
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10107-022-01801-y
Keywords
- Submodular function maximization
- Sequence-independent lifting
- Polyhedra
- Multidimentional lifting
- Mixed 0-1 knapsack