Abstract
Linear bilevel optimization problems are often tackled by replacing the linear lowerlevel problem with its Karush–Kuhn–Tucker conditions. The resulting singlelevel problem can be solved in a branchandbound fashion by branching on the complementarity constraints of the lowerlevel problem’s optimality conditions. While in mixedinteger singlelevel optimization branchandcut has proven to be a powerful extension of branchandbound, in linear bilevel optimization not too many bileveltailored valid inequalities exist. In this paper, we briefly review existing cuts for linear bilevel problems and introduce a new valid inequality that exploits the strong duality condition of the lower level. We further discuss strengthened variants of the inequality that can be derived from McCormick envelopes. In a computational study, we show that the new valid inequalities can help to close the optimality gap very effectively on a large test set of linear bilevel instances.
The difficulty in closing the optimality gap
Roughly speaking, branchandbound algorithms solve mathematical optimization problems by successively finding lower and upper bounds on the optimal objective function value. This procedure progressively decreases the optimality gap, i.e., the difference of the two bounds, until it is closed and the lower and upper bound meet. For minimization problems, every primal feasible solution provides a valid upper bound on the objective function value. Lower bounds in turn are computed by solving relaxations of the original problem. While modern branchandbound algorithms may find good primal solutions quickly, proving optimality by closing the optimality gap might be very challenging. It is not unusual to observe solution processes similar to the dashed line in Fig. 1, which shows an exemplary evolution of the lower and upper bounds over the number of visited nodes provided by a branchandbound implementation. An almost optimal solution is found right at the beginning, but the lower bound improves only slowly. As a result, many branchandbound nodes need to be visited until the gap is closed and optimality is proved.
In mixedinteger programming, the discussed obstacle has been tackled by subsequently adding valid inequalities that cut off integerinfeasible points. In many cases, this yields tighter relaxations and ultimately delivers stronger lower bounds. Such branchandcut algorithms are now stateoftheart in solving mixedinteger problems.
Linear bilevel problems, in which some variables of a linear upperlevel problem need to constitute an optimal solution of a second linear optimization problem (the lowerlevel problem), are no exception to the behavior discussed above in general. While bilevelfeasible points, i.e., points that satisfy all upperlevel constraints and lowerlevel optimality, can often be found quickly [17], proving optimality is much more difficult. In fact, the dashed lines in Fig. 1 is based on a simple branchandbound code for linear bilevel problems applied to an exemplary instance. Similarly to mixedinteger programming, valid inequalities could be used to provide tighter relaxations of bilevel problems by cutting off bilevelinfeasible points, i.e., points that violate optimality of the lowerlevel problem. However, for linear bilevel problems not many tailored valid inequalities are known.
In this paper, we derive such a valid inequality for linear bilevel problems by exploiting the strongduality condition of the lowerlevel problem. This primaldual inequality turns out to be very effective for some instances. Indeed, applying it to the same instance that was used for the dashed plot in Fig. 1 yields much faster convergence; see the solid plot in Fig. 1. The lower bound increases much quicker, which results in around 20 000 visited nodes compared to roughly 45 000 nodes when the inequality is not used. We will analyze the benefit gained by the proposed valid inequality in detail in a computational study later in the paper.
The remainder of the paper is structured as follows. In Sect. 2 we formally introduce linear bilevel problems and review existing valid inequalities. Afterward, we develop a new valid inequality based on the strongduality condition of the lowerlevel problem in Sect. 3 and also propose some tighter variants. In Sect. 4, we evaluate the effectiveness of the inequalities in a computational study. Finally, we conclude in Sect. 5.
Linear bilevel problems and valid inequalities
In this paper, we consider linear bilevel problems of the form
where \(\mathcal {S}(x)\) denotes the set of optimal solutions of the parameterized linear program
with \(c \in \mathbb {R}^n\), \(d, f \in \mathbb {R}^m\), \(A \in \mathbb {R}^{k \times n}\), \(B \in \mathbb {R}^{k \times m}\), \(a \in \mathbb {R}^k\), \(C \in \mathbb {R}^{\ell \times n}\), \(D \in \mathbb {R}^{\ell \times m}\), and \(b \in \mathbb {R}^\ell \). The upperlevel player (or leader) optimizes the upperlevel problem (1) by anticipating the optimal reaction \(y\) of the lowerlevel player (or follower). Whenever the follower is indifferent for a given \(x\), the set of optimal solutions \(\mathcal {S}(x)\) is not a singleton. In this case, the formulation in (1) establishes the socalled optimistic solution, i.e., the leader may select any solution \(y\in \mathcal {S}(x)\) that is the most favorable one for the upperlevel problem; see [5]. Furthermore, throughout the paper, we make the following standard assumption (see, e.g., [1,2,3]) that is necessary in Sect. 3 for the derivation of a valid inequality for Problem (1).
Assumption 1
The shared constraint set
is nonempty and bounded.
In general, bilevel problems are intrinsically nonconvex due to their hierarchical structure and even linear bilevel problems are known to be strongly NPhard [14]. In addition, even checking local optimality is NPhard; see [23]. For many realworld problems that require a bilevel or even multilevel modeling, applicationspecific solution techniques have been developed. This includes but is not limited to fields such as energy markets [8, 13, 15], pricing problems [18, 19], or network interdiction problems [4, 10]. In a more general setting in which no problemspecific structure can be exploited, most solution techniques resort to an equivalent singlelevel reformulation. For linear bilevel problems, this is typically done by replacing the lowerlevel problem (2) by its necessary and sufficient Karush–Kuhn–Tucker (KKT) conditions, which yields a mathematical program with complementarity constraints:
This reformulation was first mentioned in [12], which also contains two solution approaches exploiting the disjunctive nature of the complementarity constraints (3d). The first one is a mixedinteger linear reformulation of the KKT complementarity constraints, which requires additional binary variables and sufficiently large bigM constants. The problem can then be solved by standard mixedinteger solvers. However, bigMs that are chosen too small can yield suboptimal or infeasible solutions [21] and verifying the correctness of a bigM constant is as hard as solving the original bilevel problem; see [16]. From today’s point of view, this method should only be used if correct bigMs can be obtained via problemspecific knowledge. The second approach mentioned in [12] overcomes this obstacle by branching directly on the complementarity constraints: for all \(j = 1, \dotsc , \ell \), either the primal lowerlevel constraint is binding, i.e., \((b  Cx Dy)_j = 0\), or \(\lambda _j = 0\) holds. This approach is evaluated in more detail in [3] and improving branching rules have been proposed in [14].
One drawback of this complementaritybased branchandbound approach (as well as of the mixedinteger approach using bigMs) is a weak root relaxation. The problem that is solved in the root node is Problem (3) without the complementarity constraints (3d). In this setting, dual feasibility of the lower level (3c) is completely decoupled from the primal upper and lowerlevel constraints (3b). In the original problem (3), these two sets of constraints are solely coupled by the complementarity constraints (3d)—the exact same constraints are initially relaxed and branched on in a bilevel branchandbound algorithm. In this view, the coupling is brought back subsequently via branching. It is thus desirable to extend such bilevel branchandbound approaches to branchandcut algorithms by adding cuts that resolve the missing coupling, either already at the root node or later in the branchandbound tree. However, up to now, not too many bilevelspecific valid inequalities are known.
In [1], the complementarity conditions (3d) are used to derive disjunctive cuts that can be applied to the root node problem. For each violated complementarity constraint, solving a linear optimization problem (LP) yields such a cut. In a very small example, the usefulness of the cut is demonstrated. It is also shown that sometimes this cut couples constraints (3b) and (3c) and sometimes it does not.
In [2], three root node cuts are presented that can be derived from the solution of the root node problem. The first one is a Gomorylike cut. For each violated complementarity constraint of the lower level, two inequalities can be derived. One of them is acting on the primal upper and lowerlevel variables and the other one on the dual lowerlevel variables. At least one of the two inequalities must be valid and is actually a cut. Since the valid one is not known, both inequalities are added to the problem and a binary switching variable is used to select the valid inequality. In this light, the two inequalities add a rather implicit coupling of the constraints (3b) and (3c). Another variant are socalled extended cuts that, similar to the Gomorylike cuts, also involve binary switching variables. However, it is noted that these cuts are deeper than the Gomorylike cuts. One can also derive two cuts that do not involve a switching variable. These cuts are called simple cuts in [2]. Again, the combination of both cuts implicitly couples the primal upper as well as lower level with the dual lower level. In a small numerical study it is shown that applying a cut generation phase at the root node that adds cuts of either one of the three types outperforms pure branchandbound.
To the best of our knowledge no other generalpurpose valid inequalities dedicated to linear bilevel problems have been published so far.
A new valid primaldual inequality
All cuts reviewed in the last section have in common that they exploit the explicit disjunctive structure of the complementarity conditions. They are all derived from a single violated complementarity condition and it is not clear which violated one should be chosen to separate a cut. In this section, we derive a valid inequality for Problem (1) based on the aggregated complementarity conditions (3d). Using dual feasibility (3c), we can substitute \(\lambda ^\top D\) with f in (3d) to obtain
This is exactly the strongduality condition of the lowerlevel problem (2), as shown in the following. For a fixed upperlevel decision x, the dual to the lowerlevel problem (2) is given by
For every primaldual feasible point \((y,\lambda )\), weak duality
holds. Thus, every primaldual feasible point satisfying Inequality (4) fulfills the strongduality equation and is primaldual optimal for the lower level. An alternative formulation of the singlelevel reformulation (3) can hence be obtained by replacing the KKT complementarity condition (3d) with the strong duality condition (4). The main drawback of this approach is the bilinear term \(\lambda ^\top Cx\) of primal upperlevel and dual lowerlevel variables. When considering only integer linking variables, as, e.g., in [25], linearizations can be applied yielding mixedinteger linear reformulations. Here, however, we study purely continuous bilevel problems. Thus, this bilinear term cannot be reformulated in a mixedinteger linear way as opposed to the KKT complementarity condition (3d).
Still, the strong duality inequality can be used to derive a valid inequality for Problem (3). A straightforward idea is to relax the nonconvex term \(\lambda ^\top Cx\) by replacing each term \(C_{i\cdot } x\) in (4) with an upper bound \(C_i^+ \ge C_{i\cdot } x\), where \(C_{i\cdot }\) denotes the ith row of C. This yields the inequality
where \(C^+\) denotes the vector of upper bounds \(C_i^+\). The rationale behind this inequality is very simple and the inequality is obviously valid. Despite, or even because of its simplicity, this inequality can be very useful. It explicitly couples the primal lowerlevel variable \(y\) to the dual lowerlevel variable \(\lambda \)—a coupling that is missing in the root node problem of branchandbound approaches. The bounds \(C_i^+\) can be obtained, e.g., from variable bounds on x. While this approach is cheap from a computational point of view, it may result in weak inequalities depending on the tightness of the bounds on x. Stronger bounds \(C_i^+\) can be computed with the auxiliary LPs
where \(\mathcal {C}\) is a constraint set containing already added valid inequalities of type (6) and might be empty. This problem is bounded due to Assumption 1, such that finite bounds \(C_i^+\) exist. In addition to the root node, Inequality (6) can also be added at any node u deeper in the branchandbound tree, where the bound \(C_i^+\) is potentially tighter due to branching or previously added inequalities of type (6). This yields tighter inequalities that are locally valid for the subtree rooted at node u. Besides already added (locally) valid inequalities, the set \(\mathcal {C}\) then also contains branching decisions, and \(\mathcal {C}\) and \(C_i^+\) in (7) both depend on the current branchandbound node u. For the ease of presentation, we omit an index u for \(\mathcal {C}\) and \(C_i^+\), because this dependence will always be clear from the context. We discuss implementation details such as the timing of the generation of valid inequalities (6) or the derivation of the bounds \(C_i^+\) in Sect. 4, where we also demonstrate the effectiveness of the inequalities in a numerical study.
Before, let us emphasize that Inequality (6) can also be derived from another perspective. Consider a general bilinear term \(z=vw\) with bounds \(v^ \le v \le v^+\) and \(w^ \le w \le w^+\). Then, McCormick envelopes [20] provide linear under and overestimators for \(z=vw\):
This can be applied to the strongduality condition (4). We can decompose the bilinear products \(\lambda ^\top C x= \sum _{i=1}^{\ell } z_i\) to obtain terms \(z_i = v_i w_i\) with \(v_i = \lambda _i\) and \(w_i = C_{i\cdot } x\). Due to the sign in the strongduality condition (4), only the overestimators (8b) can be used:
If we apply the initial bounds \(\lambda _i^ = 0\) for all \(i=1,\ldots ,\ell \), then (9b) simplifies to
Obviously, Inequality (6) is fulfilled if (9a) and (10) are satisfied. Contrary, when Inequality (6) is feasible, then \(z_i = \lambda _i C_i^+\) is feasible for (9a) and (10) Thus, (9a) together with (10) is equivalent to Inequality (6). However, whenever tighter (local) bounds \(\lambda _i^ > 0\) are available, e.g., after presolve or branching, (9a) and (9b) provide a tightening of (6). The second overestimator (9c) involves bounds \(C_i^\le C_{i\cdot } x\), which can again be obtained by variable bounds on \(x\) or by minimizing instead of maximizing in Problem (7). However, it also involves upper bounds \(\lambda _i^+\) for the initially unbounded dual variables \(\lambda _i\). In general, such dual upper bounds are not available so that the overestimator (9c) cannot be used. Yet, whenever a (maybe locally valid) bound for \(\lambda _i\) is available by chance, e.g., due to a combination of branching and node presolve, the overestimator (9c) can be used to potentially tighten the valid inequality (6). In this light, the derivation via McCormick envelopes (8) may indeed provide tighter versions of Inequality (6). While the applicability of the tighter variants of the inequality solely depends on the availability of bounds, the basic inequality (6) can always be derived. We will discuss the applicability of the tightened variants in Sect. 4.
Furthermore, one could also relax \(\lambda ^\top Cx\) in the strongduality inequality (4) by replacing each term \(\lambda ^\top C_{\cdot j}\) with an upper bound \(C_j^+\ge \lambda ^\top C_{\cdot j}\), where \(C_{\cdot j}\) denotes the jth column of C. We then obtain the inequality
This inequality couples all three types of variables \(x\), \(y\), and \(\lambda \) and can also be derived from the McCormick envelopes (8) by decomposing \(\lambda ^\top C x= \sum _{j=1}^{n} z_j\) with \(z_j = v_j w_j\), \(v_j = \lambda ^\top C_{\cdot j}\), and \(w_j = x_j\). However, Inequality (11), respectively both overestimators (8b), involve finding lower or upper bounds \(C_j^\pm \) for \(\lambda ^\top C_{\cdot j}\). This means that every problem
needs to be bounded to obtain finite coefficients for each \(x_j\). The lowerlevel problem (2) is bounded due to Assumption 1. Thus, the feasible set \(\Omega _D\) of the dual lowerlevel problem (5) is bounded in the direction \(bCx\) of the dual objective function. However, this is not necessarily the case for the optimization directions \(C_{\cdot j}\). In fact, preliminary computational tests revealed that no instance in our test set has the property that all problems (12) are bounded. We thus refrain from using Inequality (11) and its variants that can be derived by McCormick envelopes. Finally, note that (6) and (11) are also valid for the pessimistic version of the bilevel problem,
since the lowerlevel problem is still given by (2). However, in order to streamline the presentation, we will stick to the discussion of the optimistic case.
Computational study
We now evaluate the effectiveness of the valid inequalities derived in Sect. 3 within a complementaritybased branchandbound framework similar to what is described in Sect. 2. All our experiments are carried out on a single thread using the C interface of CPLEX 12.10 on a compute cluster with Xeon E31240 v6 CPUs at 3.7 GHz and 32 GB RAM; see [22] for more details.
Our complementaritybased branchandbound algorithm is realized in the following way. We introduce slack variables \(s_i = b_i  C_{i\cdot } x D_{i\cdot } y\ge 0\) to the singlelevel reformulation (3) for every lowerlevel constraint. We can then rewrite the complementarity constraints (3d) using specialorderedsets of type 1 (SOS1) for each pair \((s_i,\lambda _i)\). This way, we could use the SOS1 capabilities of CPLEX to branch on the complementarity conditions. However, to have full control and information on the branching (in particular, on the set \(\mathcal {C}\)), we implemented our own branching and bookkeeping using generic CPLEX callbacks. We branch on the most violated complementarity constraint \(i \in \{1, \dotsc , \ell \}\) by setting either \(s_i = 0\) or \(\lambda _i = 0\), while leaving the node selection to CPLEX. This basic branchandbound procedure serves as a benchmark and is called B&B throughout this section. Interestingly, a preliminary computational study revealed that B&B already outperforms the native SOS1 branching of CPLEX.
We extend this setting to a branchandcut approach by subsequently adding the valid inequalities described in Sect. 3 via generic CPLEX callbacks. We therefore use the general formulation (9). This allows to add tighter inequalities whenever the required bounds are available. In a preliminary computational study, we tested various inequalities and strategies of how and when to add the inequalities. It turned out that computing the bounds \(C_i^\pm \) and \(\lambda _i^\pm \) with auxiliary LPs, similar to Problem (7), provides significantly better bounds and thus tighter inequalities than using internal global and local bounds provided “for free” by CPLEX. Although timeconsuming, we follow the former approach to generate the tightest inequalities possible. Our preliminary experiments also revealed that making use of the McCormick overestimators (9b) and (9c) by tightening \(\lambda _i^\) and \(C_i^\) is only beneficial for a very small fraction of tested instances and in most cases it even harms the solution process. Hence, in the remainder of this section, we only discuss results for Inequality (6), implemented as the set of inequalities (9a) and (10). In particular, we compare the following parameterizations, where \(\ell \in \mathbb {N}\) denotes the number of lowerlevel primal constraints:
 B&B:

: The branchandbound benchmark without additional inequalities.
 C&B:

: The set of inequalities (9a) and (10) is added at the root node if violated.
 B&C(5):

: Inequality (9a) is added at the root and the inequalities (10) are added whenever (6) is violated at a node with depth \(d = p \lfloor \ell /5\rfloor \), \(0 \le p \in \mathbb {N}\).
 B&C(10):

: Like B&C(5) but with \(d = p \lfloor \ell /10\rfloor \).
Obviously, the separation routine is invoked twice as many times in B&C(10) compared to B&C(5).
To compare our different methods, we use linear bilevel instances described in [17]. Table 1 summarizes the sizes of different test sets. The column “reference” indicates the origin in the literature of each subset and in the column “total” we state the size of the respective test sets. Further, the column “solved” shows how many instances are solved by at least one of the above methods in a time limit of 1 h, whereas “easy” indicates how many are solved in less then 10 s by all four methods. Finally, the last column displays the remaining number instances for each test set. Note that the test set XU consists of the test sets XUWANG and XULARGE, which are constructed the same way. Furthermore, based on our preliminary computational experiments, we completely omit the test sets DENEGRE, GENERALIZED, as well as INT0SUM since they are too easy (i.e., all instances are labeled “easy”) and GK, INTERFIRE, as well as MIPLIB since they are too hard (i.e., hardly any instance is labeled “solved”). We thus obtain a total of 408 instances in Table 1. In the following, we discuss our observations w.r.t. the remaining instances in each of these different test sets. We illustrate the performance of the different parameterizations of our implementation using performance profiles according to [7]. For each instance \(i\) and implementation variant \(s\), we compute the performance ratio
w.r.t. the branchandbound node count, where \(S\) is the set of all studied implementation variants. This means that \(n_{i,s}\) is the node count of variant \(s\) on instance \(i\). Every performance profile for node counts in this section shows the proportion of instances for which a given approach lies within a factor \(\tau ^\text {n}\ge 1\) of the best approach. Similarly, we introduce \(\tau ^\text {t}\) for performance profiles w.r.t. the running times in wallclock seconds.
It is well known that cuts often work only on a small number of instances and not throughout large and diverse test sets, in particular if they exploit a certain structure. Thus, we first discuss the impact of the valid inequalities for specific subsets of instances. It has already been shown in Fig. 1 that the application of our valid inequalities is capable of closing the optimality gap much faster compared to a pure branchandbound. This effect is even more pronounced for all instances of the test set CLIQUE. These instances are solved immediately once the valid inequality is added at the root node. In contrast, B&B finds the optimal solution early in the tree in most of the cases but the lower bound does not improve at all. Thus, B&B cannot solve a single instance within the time limit of 1 h. For INTERCLIQUE, we observe a similar behavior, except that a few instances can also be solved by B&B.
On the other hand, for the test set KP, it is beneficial to also separate inequalities further down in the branchandbound tree. Figure 2 shows performance profiles for branchandbound node counts (left) and total running times (right) for these instances. We first discuss the node counts and observe that C&B yields a notable improvement over B&B. However, C&B in turn is clearly dominated by B&C(10), which needs the least branchandbound nodes for almost every instance. On the other hand, this comes at a certain price since the node count improvement is not significant enough to compensate the time needed to separate the additional cuts; see also the right plot in Fig. 2. Thus, C&B yields the best performance in terms of running times and dominates every other approach. The results on the test set INTERASSIG show similar trends w.r.t. nodes, but in contrast to KP, B&C(10) is also the best performing variant in terms of running times.
While similar trends can also be observed for the node counts for the test sets INTERKP and IMKP, the decrease in nodes is insufficient to justify a branchandcut framework. In other words, B&B is dominated by every other approach in terms of node counts, but the resulting gain in running time is outweighed by cut separation, such that B&B slightly dominates the other variants in terms of running times.
Figure 3 displays performance profiles for nodes and running times restricted to the XU instances. Here, all variants perform pretty similar with respect to the node count. Since cut generation always costs computational time, it is not beneficial regarding running time to use the additional valid inequalities at all. This is especially notable for larger instances with many variables for which a large number of LPs (7) need to be solved to compute the coefficients of the cuts.
Overall, our methods are very useful on the considered instances. Figure 4 shows performance profiles for node counts and running times aggregated for all 408 instances. The branchandcut variants solve roughly 30 % more instances than the plain branchandbound procedure. All branchandcut variants largely outperform B&B, but there is no significant difference between the variants of the branchandcut method—neither in terms of node counts nor in terms of running times. To sum up, the C&B approach seems to be the best choice in general but the structure of specific instances might also lead to improved numerical results if the inequalities are added further down in the branchandbound tree.
Conclusion
In this paper, we derived a new valid primaldual inequality for linear bilevel problems based on the strongduality condition of the linear lowerlevel problem. We further discussed tightened variants of the inequality resulting from McCormick envelopes and tested these inequalities in a computational study. While the latter inequalities are not beneficial in practice, the former simple variant is shown to be crucial for proving optimality for the majority of all tested instances. In fact, for many instances, adding a single inequality at the root node is sufficient to immediately close the optimality gap. For other instances, it is shown to be beneficial to add the inequality in a branchandcut approach further down in the branchandbound tree. Overall, adding the proposed valid inequalities helps to close the optimality gap much faster compared to a pure branchandbound algorithm and gives rise to a dedicated branchandcut implementation for linear bilevel problems.
While being out of scope of this short paper, we see several enhancements that could be applied within a sophisticated branchandcut implementation for linear bilevel problems. First, adding initial valid inequalities already before preprocessing could further improve node counts and running times. Second, in case that the inequality added in the root node does not immediately prove optimality, applying several rounds of adding valid inequalities and bound tightening could be useful. Third, whenever the separation of our inequalities yields bounds \(\lambda _i^ > 0\), one could directly fix the corresponding primal lowerlevel constraint to be active. Finally, although our implemented branching rule already outperforms the SOS1based branching of CPLEX, other branching and node selection rules may further improve the performance of the overall branchandcut implementation.
References
Audet, C., Haddad, J., Savard, G.: Disjunctive cuts for continuous linear bilevel programming. Optim. Lett. 1(3), 259–267 (2007). https://doi.org/10.1007/s1159000600243
Audet, C., Savard, G., Zghal, W.: New branchandcut algorithm for bilevel linear programming. J. Optim. Theory Appl. 134(2), 353–370 (2007). https://doi.org/10.1007/s1095700792634
Bard, J.F., Moore, J.T.: A branch and bound algorithm for the bilevel programming problem. SIAM J. Sci. Stat. Comput. 11(2), 281–292 (1990). https://doi.org/10.1137/0911017
Caprara, A., Carvalho, M., Lodi, A., Woeginger, G.J.: Bilevel knapsack with interdiction constraints. INFORMS J. Comput. 28(2), 319–333 (2016). https://doi.org/10.1287/ijoc.2015.0676
Dempe, S.: Foundations of Bilevel Programming. Springer, Berlin (2002). https://doi.org/10.1007/b101970
DeNegre, S.: Interdiction and discrete bilevel linear programming. Ph.D. thesis. Lehigh University, (2011). https://preserve.lehigh.edu/etd/1226
Dolan, E.D., Moré, J.J.: Benchmarking optimization software with performance profiles. Math. Program. 91(2), 201–213 (2002). https://doi.org/10.1007/s101070100263
Egerer, J., Grimm, V., Kleinert, T., Schmidt, M., Zottl, G.: The Impact of Neighboring Markets on Renewable Locations, Transmission Expansion, and Generation Investment. Eur J Ope Res (2020) (forthcoming)
Fischetti, M., Ljubić, I., Monaci, M., Sinnl, M.: A new generalpurpose algorithm for mixedinteger bilevel linear programs. Oper. Res. 65(6), 1615–1637 (2017). https://doi.org/10.1287/opre.2017.1650
Fischetti, M., Ljubić, I., Monaci, M., Sinnl, M.: Interdiction games and monotonicity, with application to knapsack problems. INFORMS J. Comput. 31(2), 390–410 (2019). https://doi.org/10.1287/ijoc.2018.0831
Fischetti, M., Monaci, M., Sinnl, M.: A dynamic reformulation heuristic for generalized interdiction problems. Eur. J. Oper. Res. 267(1), 40–51 (2018). https://doi.org/10.1016/j.ejor.2017.11.043
FortunyAmat, J., McCarl, B.: A representation and economic interpretation of a twolevel programming problem. J. Oper. Res. Soc. 32(9), 783–792 (1981). https://doi.org/10.1057/jors.1981.156
Grimm, V., Schewe, L., Schmidt, M., Zöttl, G.: A multilevel model of the european entryexit gas market. Math. Methods Oper. Res. 89(2), 223–255 (2019). https://doi.org/10.1007/s001860180647z
Hansen, P., Jaumard, B., Savard, G.: New branchandbound rules for linear bilevel programming. SIAM J. Sci. Stat. Comput. 13(5), 1194–1217 (1992). https://doi.org/10.1137/0913069
Hu, X., Ralph, D.: Using EPECs to model bilevel games in restructured electricity markets with locational prices. Oper. Res. 55(5), 809–827 (2007). https://doi.org/10.1287/opre.1070.0431
Kleinert, T., Labbé, M., Plein, F., Schmidt, M.: There’s No Free Lunch: On the Hardness of Choosing a Correct BigM in Bilevel Optimization. Oper Res 68(6), 1716–1721 (2020). https://doi.org/10.1287/opre.2019.1944
Kleinert, T., Schmidt, M.: Computing stationary points of bilevel problems with a penalty alternating direction method. INFORMS J. Comput. (2019). https://doi.org/10.1287/ijoc.2019.0945
Labbé, M., Marcotte, P., Savard, G.: A bilevel model of taxation and its application to optimal highway pricing. Manag. Sci. 44(12 part 1), 1608–1622 (1998). https://doi.org/10.1287/mnsc.44.12.1608
Labbé, M., Violin, A.: Bilevel programming and price setting problems. 4OR 11(1), 1–30 (2013). https://doi.org/10.1007/s1028801202130
McCormick, G.P.: Computability of global solutions to factorable nonconvex programs: Part IConvex underestimating problems. Math. Program. 10(1), 147–175 (1976). https://doi.org/10.1007/2FBF01580665
Pineda, S., Morales, J.M.: Solving linear bilevel problems using bigms: not all that glitters is gold. IEEE Trans. Power Syst. (2019). https://doi.org/10.1109/TPWRS.2019.2892607
Regionales Rechenzentrum Erlangen. Woodcrest Cluster. url: https://www.anleitungen.rrze.fau.de/hpc/woodycluster (visited on 06/03/2020)
Vicente, L., Savard, G., Júdice, J.: Descent approaches for quadratic bilevel programming. J. Optim. Theory Appl. 81(2), 379–399 (1994). https://doi.org/10.1007/BF02191670
Xu, P., Wang, L.: An exact algorithm for the bilevel mixed integer linear programming problem under three simplifying assumptions. Comput. Oper. Res. 41, 309–318 (2014). https://doi.org/10.1016/j.cor.2013.07.016
Zare, M.H., Borrero, J.S., Zeng, B., Prokopyev, O.A.: A note on linearized reformulations for a class of bilevel linear integer problems. Ann. Oper. Res. 272(1), 99–117 (2019). https://doi.org/10.1007/s104790172694x
Acknowledgements
This research has been performed as part of the Energie Campus Nürnberg and is supported by funding of the Bavarian State Government. The authors thank the DFG for their support within project A05, B08, and Z01 in CRC TRR 154. Fränk Plein thanks the “Fonds de la Recherche Scientifique” (F.R.S.FNRS) for financial support. Martine Labbé has been partially supported by the Fonds de la Recherche Scientifique  FNRS under Grant(s) no PDR T0098.18.
Funding
Open Access funding enabled and organized by Projekt DEAL.
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.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Kleinert, T., Labbé, M., Plein, F. et al. Closing the gap in linear bilevel optimization: a new valid primaldual inequality. Optim Lett 15, 1027–1040 (2021). https://doi.org/10.1007/s11590020016606
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11590020016606
Keywords
 Bilevel optimization
 Valid inequalities
 Branchandcut
 Computational analysis
Mathematics Subject Classification
 90Cxx
 9008
 90C11
 90C46