Mixed-integer optimal control under minimum dwell time constraints

Tailored Mixed-Integer Optimal Control policies for real-world applications usually have to avoid very short successive changes of the active integer control. Minimum dwell time (MDT) constraints express this requirement and can be included into the combinatorial integral approximation decomposition, which solves mixed-integer optimal control problems (MIOCPs) to ϵ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon $$\end{document}-optimality by solving one continuous nonlinear program and one mixed-integer linear program (MILP). Within this work, we analyze the integrality gap of MIOCPs under MDT constraints by providing tight upper bounds on the MILP subproblem. We suggest different rounding schemes for constructing MDT feasible control solutions, e.g., we propose a modification of Sum Up Rounding. A numerical study supplements the theoretical results and compares objective values of integer feasible and relaxed solutions.


Introduction
We consider mixed-integer optimal control problems (MIOCPs) on the fixed and finite time horizon T := [t 0 , t f ] ⊂ R of the following form inf for t ∈ T , (1.1) (1. 3) The differential states x ∈ W 1,∞ (T , R n x ) with fixed initial values x 0 ∈ R n x are governed by the right-hand side ordinary differential equation (ODE) function f : R n x × {v 1 , . . . , v n ω } → R n x , which is assumed to be continuous in the first argument. We further assume that there exists a solution x for the above problem. Let V = L ∞ (T , {v 1 , . . . , v n ω }) and {v 1 , . . . , v n ω } ⊂ R n v so that the discrete control function v : T → {v 1 , . . . , v n ω } is assumed to be measurable and is further restricted by minimum dwell time (MDT) constraints represented by the subset V MDT . We stress that this function takes values out of a finite set with cardinality n ω ∈ N and exclude the trivial case n ω = 1. We minimize Φ ∈ C 0 (R n x , R) over the end state, which in turn depends on the discrete control function v.

Related work
Optimal control problems with integer control choices have been investigated in many research articles during the last few decades [12,14,31,40] since they naturally arise in a range of applications. In order to apply the obtained control policy in practice, additional switching constraints are usually needed, such as minimum dwell time requirements that describe the necessity of activating an integer control for at least a given minimal duration if at all and analogously for deactivation periods. Recent case studies of MIOCPs with MDT considerations can be found, e.g., for pesticide scheduling in agriculture [2], electric transmission lines [17], solar thermal climate systems [8] and hybrid electric vehicles [30]. MDT constraints attracted also a lot of attention as part of mixed-integer linear programs (MILPs), see [29] for a study of unit-commitment problems and [25] for a corresponding polytope investigation. For a recent work about model predictive control under MDT constraints see [10].
There are MIOCP related problems and approaches that have been discussed in the literature. For example, the optimal control community has been successfully solving so-called bang-bang problems for decades. In contrast to the above problem formulation, the considered linear control problems are required to have the bang-bang property, i.e., the derivative of the Hamiltonian is strictly positive or strictly negative almost everywhere. The main challenge consists in guessing the correct switching order and numerically detecting the switching points. This approach does not work for problems that involve singular or path-constrained arcs. This can be overcome by using a discrete global maximum principle (see [15] for further references). Still, the indirect first-optimize-then-discretize approaches have some drawbacks compared to direct methods and it is not clear how combinatorial constraints such as MDT constraints could be incorporated. As part of direct methods, one common approach to tackle MIOCPs under MDT constraints is to consider two separate levels of optimization. At the upper level, a mode insertion gradient is usually evaluated in order to fix a sequence of active system modes with promising cost function value. At the lower level, the algorithm aims at minimizing the cost functional with respect to the switching times and continuous control input, if available. Such approaches can be found in [1,4,12,14,40] and are usually referred to as transition-time optimization or variable time transformation method. We remark such switched systems have also been intensively investigated with respect to stability under average dwell time, see [20]. Another approach to include MDT constraints into MIOCPs is to apply dynamic programming, see [9], which is, however, computationally expensive.

Results for the CIA decomposition
Similar to the two-stage optimization method, the combinatorial integral approximation (CIA) decomposition [31,35] that uses the same approximation theory as the embedding transformation technique [5], is based on the idea to solve the problem in several optimization steps. The first idea is to use the partial outer convexification method [31,34] that allows us to reformulate a problem with integer controls into an equivalent problem with a {0, 1} n ω -valued control function ω, which indicates that exactly one discrete realization v i ∈ {v 1 , . . . , v n ω } is active for each time point. For this, we replace the Constraint (1.1) bẏ for t ∈ T . (1.4) Second, discretizing the MIOCP in the spirit of first-discretize-then-optimize methods with, e.g., Direct Multiple Shooting [6] or Direct Collocation [37] results in a mixed-integer nonlinear program (MINLP). We introduce a time discretization by the ordered set G N := {t 0 < · · · < t N = t f } denoting a grid with N intervals and lengths Δ j := t j − t j−1 ,Δ := max j Δ j as well as Δ := min j Δ j for j ∈ [N ]. The binary control functions ω i (·) are assumed be piecewise constant, changing values only on these grid points, so that ω(·) can be uniquely represented by a matrix w ∈ {0, 1} n ω ×N . This MINLP becomes a nonlinear program (NLP), if the binary controls are relaxed. We write a ∈ [0, 1] n ω ×N for denoting this relaxed value. After solving this NLP, the resulting a is approximated in the CIA step -which is an MILP -with binary control values w. The main idea of this CIA problem is to minimize the integrality gap θ(w), which is the accumulated control deviation between a and w, i.e., min w θ(w), where θ(w) := max This work is based on the CIA decomposition due to its advantages: 1. MINLPs fall generally into the class of NP-hard problems so that using approaches that bypass the direct solution of such problems is computationally favorable. 2. Convergence results have been proven for MIOCPs without MDT constraints in the sense that, under mild assumptions, the obtained solution was shown to be arbitrarily close to the optimal solution with grid lengthΔ going to zero [26,31,34]. Moreover, the solution of the NLP represents a useful a priori lower bound on the objective, if solved to global optimality. 1 3. An MILP enables the option to include intuitively a large variety of combinatorial constraints. Numerical case studies showed that the resulting feasible solution is close to the relaxed solution in case the applied combinatorial constraints are not too restrictive [7,8].
For a generalization of the CIA decomposition to more optimization steps and MILP variants see [41] as well as to the PDE constraint case see [18,19,27]. Recently, an extension of the algorithm with the penalty alternating direction method has been made [16], which can be regarded as a feasibility pump [13] variant for MIOCPs. Instead of minimizing θ(w) by means of solving an MILP, fast rounding algorithms can be applied, such as sum-up rounding (SUR) [31] and next forced rounding (NFR) [21], which generate binary solutions that still converge to the relaxed binary control solution with vanishing grid length. The SUR scheme computes for j = 1, . . . , N For defining the NFR algorithm, one needs for all i = 1, . . . , n ω and iteratively for j = 1, . . . , N the quantity else.
(1.5) A control with index i ∈ [n ω ] on interval j is defined to be next forced, if and only if Then, the NFR algorithm sets iteratively for j = 1, . . . , N the next forced control equal to one (break ties arbitrarily) and if there is no such control, the active control is chosen according to the SUR scheme. We summarize established integrality gap bounds for these schemes in the form of θ(w) ≤ C(n ω )Δ in Table 1. Table 1 So far known integrality gap bounds for binary control approximation algorithms Next-forced rounding (NFR) Sum-up rounding (SUR) SUR with vanishing constraints C(n ω ) = 1, see [21], n ω i=2 1 i , see [24], n ω /2 , see [28] By vanishing constraints we refer to constraints of the form ω i (t)g(x(t)) ≤ 0, for i = 1, . . . , n ω where g is a smooth function

Contribution
We conduct a theoretical analysis of the integrality gap in the presence of MDT constraints. In particular, we prove the upper bound where C U D denotes the maximum of the minimum up (MU) and minimum down (MD) duration of all controls. We show that this bound is tight for MU time constraints. As a consequence of this result, the tightest possible bound for the CIA problem is The proof is constructive, as we introduce a generalization of the NFR scheme to the MDT setting. We present a rounding modification so that MD times can be included explicitly allowing us to deduce an improved integrality gap bound for this case. In the same spirit, we modify SUR so that MDT requirements are satisfied by the obtained binary solution. We test our algorithms on the three tank problem from a benchmark library [33] and evaluate with increasingly restrictive MDT constraints how large the gap between the constructed integer feasible and relaxed solution becomes.

Outline
We give a problem definition of the MIOCP in Sect. 2 and describe the proposed CIA decomposition algorithm with (CIA) as subproblem in detail in Sect. 3. A generalization of NFR to the MDT setting is presented in Sect. 4. It provides the tools to derive upper bounds on the integrality gap in Sect. 5. The commonly used SUR scheme can be extended to include also MDT constraints, which we present in Sect. 6 together with a discussion of the integrality gap. Section 7 provides a numerical example and we finish this article with conclusions in Sect. 8.

Mixed-integer optimal control problem
We introduce our problem at hand in the discretized and (partial) outer convexified setting, but refer to [22,31] and, more recently, to [23,24] for extensive descriptions of partial outer convexification, the continuous MIOCP setting and its relation to the discretized problem.

Definition of binary and relaxed controls
We base the following definitions on the grid G N we already introduced in the previous section.

Definition 1 (Convex combination constraint (Conv) and Matrix sets W , A)
We express the requirement that the columnwise entries of a matrix (m i, j ) ∈ [0, 1] n ω ×N sum up to one by and call it convex combination constraint (Conv) in the sequel. Based on this constraint, we define the sets where we denote by Conv(W ) the convex hull of W .
We notice the geometric nature of W and A. They are the vertices respectively the set of faces of the N -fold iterated standard simplex without the origin and spanned by the n ω unit vectors. Definition 2 (Binary ω and relaxed control functions α) Let the vector of binary control functions ω and its corresponding vector of relaxed control functions α be defined by their function space domains where w ·, j , respectively a ·, j , denotes the jth column of w, respectively of a. 2

Optimal control problem class
We take interest in the discretized binary control problem (DBCP) with MDT constraints and its naturally relaxed problem, the discretized relaxed control problem (DRCP), which arises from replacing ω by α.

Definition 3 (Problems (DBCP) and (DRCP))
Let a fixed discretization grid G N be given together with an MU time C U ≥ 0 and an MD time C D ≥ 0. We refer to the following general problem class as (DBCP) (discretized) outer convexified dynamics (1.4), initial value condition for x (1.2), piecewise constant controls: ω ∈ Ω, MDT constraints: where we denote the intervals affected by the MDT We define a control i ∈ [n ω ] to be active on the interval starting with t j−1 ∈ G N , if and only if ω i (t) = 1 for t ∈ [t j−1 , t j ) is true and the other way around for inactive controls. If a binary control is active after a switch on t j , it has to stay active for a time period of at least C U as required by (2.1), whereas (2.2) enforces the analogous case for deactivating a control. Finally, we define (DRCP) as the canonical relaxation of problem (DBCP) where we optimize over α ∈ A instead of ω ∈ Ω.
We remark that our study assumes no mode specific MU times C i,U or MD times C i,D , but may include them by setting C U = max i∈[n ω ] C i,U and accordingly C D = max i∈[n ω ] C i,D , even though this simplification may result in suboptimal solutions.
Without loss of generality, we omit in our problem definition further constraints and continuous controls u ∈ L ∞ (T , R n u ). See [32] for a discussion of extensions and [27] for PDE constraints.

Combinatorial integral approximation decomposition
We recapitulate the CIA decomposition algorithm to solve MIOCPs in Fig. 1. This approach is justified by a convergence result in the situation without MDT constraints establishing that the objective integer gap of (DBCP) compared to (DRCP) depends linearly on the integrated control deviation under certain regularity assumptions [26,31].  1 Schematic representation of the CIA decomposition. MIOCPs can be equivalently reformulated into their partially outer convexified counterpart problem that is thereafter transformed into (DBCP) via a temporal discretization. Next, allowing a convex combination a in (Conv) yields the relaxed problem (DRCP). After solving this problem we obtain a , which are then approximated with binary values w in the rounding problem (CIA). Finally, we use w as fixed variables for solving (DBCP) as a continuous variable problem Thus, we are interested in an admissible ω represented by its corresponding value w ∈ W with integrality gap bounded by the grid length in the sense of as already mentioned in the introduction, so that the solution of (DRCP) can be arbitrarily well approximated by an integer solution through refining the discretization grid [34]. The SUR scheme constructs solutions ω with this property; in fact, Kirches et al. [24] showed C ∈ O(log(n ω )) as conjectured in [34]. Hence, the integrality gap obtained by this rounding scheme is going to be arbitrarily small with arbitrarily small grid length, which is why SUR is often applied in the CIA decomposition as rounding step after solving (DRCP). Rather than using SUR, it has been proposed [35] to formulate the problem (3.1) as an MILP for an improved approximation and to be able to consider also different norms (such as the Manhattan norm) and combinatorial constraints. Therefore, we recall its definition and state its MDT variants. Definition 4 (Problems (CIA), (CIA-U), (CIA-D), and (CIA-UD)) Let a ∈ A be the (local) optimal solution of (DRCP) and assumed to be given. Then, we define the problem (CIA) to be Clearly, (CIA) is a reformulation of minimizing (3.1). In Sect. 5 we are going to elaborate upper bounds for (3.1) in the presence of MDT constraints by investigating its (CIA) variants. Since the lower bound on its objective is trivially zero (and can be reached), bounds always refer to upper bounds in this article.
With respect to the constraints (2.1) and (2.2) we stress that there are other, in many cases computationally more efficient, formulations of MDT constraints, e.g. in the spirit of extended formulations [25]. The latter may lead to relaxations that are less likely to deliver fractional solutions [25] and thus can be beneficial for including the constraints (2.1) and (2.2) into the NLP solving procedure. But, as this study does not focus on the NLP formulation, i.e., how a is achieved, and since we propose to solve (CIA-UD) by means of tailored rounding heuristics or a branch and bound algorithm and not with a standard MILP solver exploiting extended formulations, we would benefit neither numerically nor theoretically from alternative MILP formulations and we therefore skip the presentation of these.

Dwell time next forced rounding
We introduced in (1.5)-(1.6) the NFR scheme as an algorithm that can compute approximations to solutions of (CIA) in O(n ω N 2 ) [21] and that constructs feasible solutions of (CIA) with objective no larger thanΔ. In this section, we introduce dwell time next forced rounding (DNFR) as a generalization, aiming for a scheme that constructs MDT constraint feasible solutions and from which we derive bounds for the (CIA) objective and its MDT variants. Several definitions are needed for DNFR and we begin with a definition of sequences of intervals that are grouped into blocks in the presence of MDT constraints.

Definition 5 (Dwell time block interval sets)
Let an MDT C 1 ≥ 0 be given. We define iteratively the dwell time invoked interval sets J b and their last indices l b for b = 1, . . . , n b and with l 0 := 0: where n b := min{b | l b = N } represents the number of interval blocks.
In the following we will sometimes write loosely block instead of dwell time block for shortening our language. Next, we establish the lengths of dwell time blocks. and name the maximum, respectively minimum, length of all dwell time blocks L , respectively L , i.e., By the definition of dwell time blocks, we see that L b depends both on the time discretization G N and on C 1 . If C 1 ≤ Δ, then, the blocks are in fact the grid intervals, i.e., L j = Δ j , j ∈ [N ] and n b = N . As soon as C 1 > Δ holds, there is at least one block b with length of two subsequent intervals Overall, one recognizes that L increases monotonically with increasing C 1 and obviously stops as soon as C 1 > t f − t 0 . The DNFR scheme relies crucially on the block dependent accumulated control deviation, which is why we introduce it as auxiliary variable in the next definition.
For controls i ∈ [n ω ] we define the accumulated control deviation on interval j ∈ [N ] as and further define θ i,0 := 0 for all i ∈ [n ω ]. We introduce for blocks b ∈ [n b ] the analogous variables In the sequel, we sometimes write forward control deviation for control i in order to The defined quantities of Definitions 5-7 are illustrated in Fig. 2.

Remark 1
If a small MDT C 1 ≤ Δ is given, θ i, j equals trivially Θ i, j and the same holds for γ i, j and Γ i, j . Nevertheless, with an MDT of C 1 > Δ one needs interval and block related variables to be able to clearly distinguish between both values.

Remark 2
Let w ∈ W and we denote by θ(w) its (CIA) objective value. With the above definition, we conclude θ(w) = max i∈[n ω ], j∈[N ] |θ i, j |. Generally, one notices that the maximum of the |θ i, j | values must be assumed at an interval before a switch happens (i.e., w ·, j = w ·, j+1 ) or on the last interval, since |θ i, j | increases monotonically with constant w ·, j and increasing j. Hence, with constant w ·, j on the dwell time blocks, we also have that We introduced in (1.5)-(1.6) the concept of a next-forced control that depends on the maximum grid lengthΔ. We generalize this concept by using blocks and a generic rounding threshold factor C 2 > 0 instead of using always C 2 = 1 as in the NFR scheme. To this end we present a definition of different types of control variable activations that depend on prior variables choices and on a ∈ A.
Definition 8 (Admissible, forced, and future forced activation) Let the rounding threshold factor C 2 > 0 and a ∈ A be given. The choice Right: corresponding accumulated control deviation. The forward control deviation with respect to intervals γ i is greater or equal to θ i with equality if the control i is inactive. Also, where the latter equals via definition the last θ i before block b begins. Notice that Γ i,2 sums up the weighted α values for the intervals 3-6, which results in a large value holds, with the special case l = b meaning i is actually forced on b. If the above inequality holds for any l ≤ n b , we call the control i ∈ W b a on block b to be future forced, and group these controls into the set W b f .

Definition 9 (Minimum down time forbidden control)
We introduce the constant χ D ∈ {0, 1}. If the CIA problem involves an MD time constraint with C D > Δ, we set . . , n ω , as the index of the control that has been activated on block b − 2 and deactivated on block b − 1 -if such a control exists: Then, let I D b denote the χ D dependent set of the minimum down time forbidden control otherwise. Note that I D b is either the empty set or contains exactly one control index. It may seem unintuitive to declare only one control per block as minimum down time forbidden, because a sufficiently large chosen MD time can comprise more than two intervals and therefore more than one control could be minimum down forbidden on certain blocks. However, in such a situation, where several controls are minimum down forbidden, it could happen that only one control may be allowed to be active, resulting in a large control deviation. Consequently, a fine granular definition would be critical for deriving bounds for (CIA-UD) using the DNFR scheme. We will specify such a worst case later in Example 1 as part of "Appendix B" and argue thereby why we tolerate at most one minimum down time forbidden control per block. We illustrate the different control activation types of Definitions 8 and 9 in Fig. 3.
Finally, we use these control properties to declare the DNFR scheme in Algorithm 1. In contrast to the original NFR, we do not iterate over all intervals, but over all dwell time C 1 invoked blocks (line 2) and check on each block whether there is a forced control and activate it in this case (line [3][4]. Otherwise, we test if there is an earliest future forced control and if so, we set it to be active (line [5][6][7][8]. Else, the algorithm selects the control with the maximum forward control deviation (line 9-13), which represents a fallback to the classical SUR scheme. In case the MD mode is turned on by setting χ D = 1, we exclude the set I D b from our control selection task (line 3, 5, 11). This consideration of minimum down time forbidden controls is a further extension of the original NFR scheme.

Algorithm 1: Dwell time next forced rounding algorithm for (CIA-UD)
Output: Feasible solution w DNFR of (CIA-UD) with approximation quality depending on Identify the control with the earliest future forced activation (break ties arbitrarily): Find the admissible control with maximum control deviation (break ties arbitrarily):

Tight bounds on the integrality gap for (CIA) with dwell time constraints
We now consider the question of how large the objective function value θ of (CIA) and its MDT variants can become. In other words we examine It turns out that the DNFR scheme is particularly suitable for deriving these integrality gap bounds. We state approximation results for (CIA) by means of DNFR constructed solutions. These results are presented as two theorems in Sect. 5.1 with specified parameter choices for C 2 and χ D . Afterwards, we deduce specific bounds for (CIA-U), (CIA-D) and (CIA-UD) and evaluate how tight they are in the Sects. 5.2-5.3. We begin this section with the trivial upper bound and another weak result in the following remark.

Remark
holds. In the right hand side, we maximize over a for given w and check, which one of the latter leads to an overall minimum objective. In this way a manipulates the control deviation to be as large as possible. That means with given w it is possible to set a i min , j = 1, j ∈ [N ], where i min is the control with smallest total accumulation j∈[N ] w i, j Δ j so that we obtain the (CIA) objective value θ = j∈ [N ] (1 − w i min , j )Δ j . With these arguments one may derive We omit the exact proof since this bound is generally weak as we will see later in this section.

Integrality gap results through dwell time next forced rounding
We examine in Theorem 1 how large the control deviation can become as part of the DNFR algorithm during an MD time phase. Based on this result we derive in Theorem 2 that DNFR constructs (CIA) feasible solutions with objective bounds depending on the rounding threshold C 2 and whether down time forbidden controls are allowed, i.e., χ D = 1. The proofs and the corresponding lemmata are moved to "Appendix A" to enhance readability for readers interested in the results and algorithms.
Theorem 1 (Control accumulation of a minimum down time forbidden control) Let a ∈ A, (C 2 , χ D ) = 3 2 , 1 and C 1 ≥ 0 be given and assume there is a minimum down time forbidden control i D ∈ I D b on block b ≥ 3 after Algorithm 1 was executed. Then, the forward control deviation satisfies Theorem 2 (Rounding gap of solution constructed by DNFR) Let a ∈ A and the following parameter settings be given: and C 1 ≥ 0. Then, w DNFR obtained by Algorithm 1 is a feasible solution of (CIA) for both cases with approximation quality Theorem 2 states only generic approximation results for (CIA) with an MDT parameter C 1 . We are going to assess the consequences for (CIA-U) by specifying C 1 and proving tightness of the resulting upper bound. Clearly, (CIA) is a special case of (CIA-U), where C U = 0, so that results for (CIA-U) are inherited by (CIA).
Proposition 1 (Upper bound for (CIA-U)) Let any MU time C U ≥ 0, grid G N and a ∈ A be given. Then, for (CIA-U) holds: Then, w DNFR is a feasible solution by Theorem 2 and by the property of DNFR to activate dwell time blocks of intervals with length at least C 1 = C U . From the definition of block lengths we conclude L < C U +Δ so that the assertion follows directly from Theorem 2.
In order to analyze the obtained bound with respect to tightness, we introduce a grid where the MDT C 1 overlaps the grid points by a small ε > 0. We determine the length of the resulting blocks in the following lemma. Definition 1 (Minimal C 1 -overlapping grid) Let us consider a non-degenerate MDT length, i.e., C 1 > 0, and let ε be C 1 ε > 0. Let further a time horizon [t 0 , t f ] be given with length at least 4C 1 , i.e., Lemma 1 (Length of blocks of a minimal C 1 -overlapping grid) The dwell time invoked blocks J b , b ∈ [n b ] of a minimal C 1 -overlapping grid as introduced in Definition 1 have the length Proof We keep in mind Definition 5 from which we deduce J 1 = {1, 2}, because t 0 + C 1 (minimally) overlaps t 1 . The next dwell time block begins at t 2 = t 0 + 2C 1 − ε and consists again of two intervals. This argumentation can be extended to the first n b − 1 blocks and by the definition of block lengths we conclude The length of the last block L n b is directly computed by the definition of N − 1 to be the last index of the grid point recursion before t f . Finally, the definition of a minimal C 1 -overlapping grid and the obtained block lengths implies With this preliminary work, we show that the deduced MU time bound is tight.
Proposition 2 (Tightness of the bound for (CIA-U)) Let an MU time C U ≥ 0 and a grid G N be given with Then, the objective bound for (CIA-U) mentioned in Proposition 1 is the tightest possible bound.
Proof Let us first consider C U > 0 and construct an example with the desired objective value by means of a minimal C 1 -overlapping grid, where we set C 1 = C U . The proposition assumes a time horizon length of at least 2C 1 (n ω − 1) so that the grid induced by Lemma 1 consists of at least n b ≥ n ω − 1 blocks. Let a ∈ A be given as Consequently, all controls i ∈ [n ω ], i = 1, assume in a the same values on each interval. After the first block, we set control i = 1 to zero, while all other variables are set to 1 n ω −1 for the remaining intervals, i.e. blocks. Next, we discuss how the optimal solution of (CIA-U) on the first n ω − 1 blocks may be chosen. Let us calculate the control deviation if we were to activate a control i = 2 . . . n ω on the first block: In the second and third equality we used Lemma 1. The values of the relaxed controls a for i = 2 . . . n ω and blocks 1, Thus, there are n ω − 1 controls with this control accumulation on n ω − 1 blocks, but activating any of these controls on the first block yields already the same control deviation. Hence, the objective of (CIA-U) with this a is at least 2n ω −3 , where ε is arbitrarily small. If we combine this result with Proposition 1, we get that 2n ω −3 2n ω −2 (C U +Δ) is the tightest possible bound. Last, we argue for the degenerate case, C U = 0, that we can create an example with length of all blocks set toΔ and obtain the same tight bound.
Corollary 1 (Tight bound on the rounding gap for (CIA)) Consider G N and a ∈ A. The objective of (CIA) is bounded by If N ≥ n ω − 1 holds, then this bound is tight.
Proof The bound follows from Proposition 1 with C U = 0 and if N ≥ n ω − 1, we are able to construct the same worst-case example as in the proof of Proposition 2, with intervals applied instead of blocks.

Implications for the Objectives of (CIA-D) and (CIA-UD)
The bound obtained for (CIA-U) can be transferred in a straightforward manner to (CIA-D) by using C 1 = C D as MDT in the DNFR scheme. However, we notice the increased number of degrees of freedom when dealing with MD times rather than MU times: only the down time forbidden control is fixed for a certain time duration in comparison with the MU time constraint situation where all controls are fixed due to the fixed active control. With this observation, we introduced in the DNFR scheme the min down mode χ D = 1 and are going to deduce in the sequel an alternative upper bound compared to the one obtained by DNFR with χ D = 0. As we are going to detect, this alternative bound is independent of n ω but not always an improvement, so that we declare the minimum of both bounds as the upper bound in the following proposition.
Proposition 3 (Bounds on the objective of (CIA-D) and (CIA-UD)) Consider any grid G N and a ∈ A. Let MU and MD times C U , C D ≥ 0 be given. Then

(CIA-D) is bounded by
Proof Generally, if C D > C U or if MU constraints are absent, we may apply the DNFR

Sum-up rounding in the Dwell time context
SUR is computationally less expensive (O(n ω N )) than the DNFR scheme executed on intervals. But, on the other hand, the last section showed DNFR constructs solutions for (CIA) with a maximum integrality gap that is less than the one of solutions constructed by SUR for n ω ≥ 3 (equality for the case n ω = 2). Since the SUR scheme is very often used for finding approximative solutions of (CIA), but does not necessarily fulfill MDT constraints, we discuss in this section a canonical extension of the algorithm to this setting.

Dwell time sum-up rounding (DSUR)
We introduce the concept of a currently activated control and dwell time blocks that depend on the initial interval and the MDT duration C 1 .

Definition 10 (Initial interval dwell time block index sets)
Let an MDT C 1 ≥ 0 be given. We define for all intervals k ∈ [N ] the initial interval dependent dwell time index sets to be Note that the above definition assumes implicitly fixed control variables for previous intervals [N ] k < j. We have I SUR 1 = ∅, because there are no down time forbidden controls on the first interval. Moreover, the set I SUR j may contain several controls, but at most n ω − 1.
Next, we give a definition of the DSUR scheme in Algorithm 2. It iterates over all intervals j ∈ [N ] and selects initially the interval representing the beginning of the time horizon, where a currently activated control does not yet exist. The control dependent MDT C i is updated in line 3 for each iteration inside the while loop so that C i equals the maximum of MU time C U and MD time C D for a currently activated control, and otherwise it is set to the MU time C U . The algorithm sets C i = C U for all controls in the first while iteration. Next, in line 4, one searches for the MD time admissible control i with maximum forward control deviation on the upcoming intervals covering the dwell time C i . If it is the currently activated one, we fix this control to be active also on the current selected interval j and increase the interval index (line 5-6). Else, the control is activated on the whole dwell time block represented by its interval indices J SUR j (C i ) and the interval index is updated accordingly (line [8][9]. Lastly, DSUR updates the set of down time forbidden controls for the next iteration in line 11. The algorithm stops as soon as the control choice has been made for the last interval N . Clearly, w DSUR is a feasible solution for (CIA), because exactly one control is active per interval. It is also feasible for (CIA-U), because whenever a currently deactivated control is activated, it dwells on active for at least the duration C U (line [8][9]. The solution also satisfies MD time constraints by the definition of I SUR j and altogether w DSUR is a feasible solution for (CIA-UD).
We transferred the main idea from the original SUR scheme to the problem setting with MDT constraints by selecting in each iteration the control with maximum forward control deviation. In the presence of MU time requirements, we need to calculate this forward accumulation for the set of next intervals with total length at least C U . For a given MD time larger than the MU time, Algorithm 2 compares the forward accumulation with length at least C D of the currently activated control with the ones of other controls with length at least C U . The idea behind this approach is to prevent a Set C i a = max{C U , C D } for the currently activated control i a , and set C i = C U for all other controls i = i a . 4 Find the control with maximum deviation (break ties arbitrarily): Set w i , j = 1 and update j = j + 1. Set w i ,l = 1, l ∈ J SUR j (C i ,1 ); 10 Update j = max{l | l ∈ J SUR j (C i ,1 )} + 1. situation in which a control gets deactivated, but is going to accumulate a large control deviation during its down time forbidden period. O(n ω N 2 ), since it sums up, on each interval and for all controls, the relaxed control values a on the next dwell time induced intervals.

Rounding gap bounds for Dwell time sum-up rounding
Kirches et al. [24] have proven the tightest possible bound on the integrality gap for the original SUR. From this we can derive implications for DSUR in the absence of MD conditions.

Theorem 3 (Tight bound for SUR integrality gap, cf. Theorem 7.1, [24]) Let w SUR be constructed from a ∈ A by means of SUR for an equidistant discretization of [t 0 , t f ] and let denote by θ(w SUR ) its (CIA) objective value. Then, the rounding gap is bounded by
which is the tightest possible upper bound.

Corollary 2 (Tight bound for DSUR integrality gap without MD times) Let C D < Δ and an MU time C U > 0 be given. Let the time horizon [t 0 , t f ] with a minimal C Uoverlapping grid be discretized and let w DSUR be constructed from a ∈ A by means of DSUR. Then, the rounding gap θ(w DSUR ) of its (CIA) objective value is bounded by
which is the tightest possible upper bound.
Proof As in the proof of Theorem 3 in [24] a dynamic programming argument can be applied, here with equidistant block length of (C U +Δ − ε) as derived in Lemma 1. With a time horizon length of n ω (C U +Δ − ε) we may analogously to the proof of Theorem 3 construct an example indicating that the bound is tight as follows.
The DSUR scheme constructs for this example a solution that switches directly after each block with length (C U +Δ − ε). Moreover, the controls i ∈ [n ω − 1] are active each on block i so that the last control n ω accumulates the asserted rounding gap until the end of block n ω − 1.

Remark 5 (DSUR as generalization of SUR)
The last corollary implicitly states that DSUR can be seen as a generalization of the original SUR algorithm, since it reduces to the latter for a negligible MDT C U , C D ≤ Δ.
Theorem 3 allows no direct conclusion for the case with absent MU times and an active MD time C D > Δ. At least, it is possible to provide worst-case examples for a in order to get a clue of how large the upper bound can be for the DSUR rounding gap without MU times. This expresses the following proposition.

Computational Experiments
We consider a three tank flow system problem with three controlling modes in order to evaluate the integrality gap in practice and to test the proposed rounding methods. It models the dynamics of an upper, middle and lower level tank, connected to each other with pipes. The goal is to minimize the deviation of certain fluid levels k 2 , k 4 in the middle, respectively lower, level tank. This problem type was discussed in a variety of publications for the optimal control of constrained switched systems [11,36] and is taken from the benchmark https://mintOC.de library [33]. The problem reads Furthermore, we add the MU and MD time constraints (2.1)-(2.2) to the three tank problem with varying C U and C D parameters. We leave open the question of the regularity of the differential states x, but we assume that there exists a unique solution that is continuous. We solve this test problem with the CIA decomposition. We applied Direct Multiple Shooting for temporal discretization with varying number of grid intervals N together with a fourth order Runge-Kutta scheme for obtaining the differential state's evolution and thus the objective value. 3 We used CasADi v3.4.5 [3] to parse the NLP with efficient derivative calculation of Jacobians and Hessians to the solver IPOPT 3.12.3 [38]. Here, the helper classes Opti stack are useful, as they allow a compact syntax for NLP modeling. For finding the optimal solution of the resulting (CIA) problem and its MDT variants we used the branch and bound solver of the open source software package pycombina 4 [8]. We published the python source code of solving (P) via the CIA decomposition online. 5 We stress that the obtained feasible solutions for (P) via the CIA decomposition are in general not global optimal solutions. In fact, the Problem (P) appears to be nonconvex so that IPOPT may construct a local solution, just like the rounding via (CIA) may do. Nevertheless, finding a global optimal solution is computational expensive as argued in the introduction. 3 When applying the fourth order Runge-Kutta scheme we need to have for the differential states that x ∈ C 5 (T , R n x ) to generate a fourth order error term. In the introduction we required only a continuous x ∈ W 1,∞ (T , R n x ); however, we could assume stronger regularity due to piecewise continuously differentiable control functions from Definition 2. Also, the Runge-Kutta scheme in the context of Direct Multiple Shooting is applied on the piecewise continuously differentiable right-hand side of the ODE, where the control function changes its values only at grid points. Nevertheless, our presented algorithms are independent of the chosen numerical integration scheme and one may choose a more accurate scheme according to the dynamical system at hand. 4 See https://github.com/adbuerger/pycombina. We illustrate in Fig. 6 the effect of an increasing MU time on the objective values of (CIA-U) and (DBCP). As expected, the finer the discretization grid and the smaller the required MDT time, the better the objective values of both problems. A small MU time results in a weak restriction for (DBCP) so that its objective value is close to the one of (DRCP), which is Φ = 8.776. But, from C U = 0.7 on, a refinement of the grid cannot compensate the MU time restriction anymore and the (DBCP) objective value is about 25% larger than (DRCP) in this case. Interestingly, this objective value increases hardly in C U ≥ 0.7, but it even decreases slightly after C U = 0.7 before increasing again and staying constant from C U ≈ 2.0 on. We observe a few outlier instances, e.g., N = 20 with C U = 1.2 or N = 40 with C U ∈ {0.4, 0.5}, where the objective value appears to be unexpectedly large. This can be explained by the coarse grid choices, which in turn stresses the importance of a fine time discretization regarding the stability of the obtained solution for (DBCP).
On the other hand, (CIA-U) objective's value increases roughly linearly in C U on fine grids until it reaches θ ≈ 0.87, which seems to be the maximum value for P in this setting. Thus, while small values of the (CIA-U)'s objective correspond to promising objective values of (DBCP), the relationship of (CIA-U) and (DBCP) appears to be quite uncorrelated from C U ≥ 0.7 on. We computed similar results for (P) with MD time constraints (not shown). We also tested whether including the relaxed MDT constraints into the NLP or not has a significant impact on the solution -this was not the case. We analyze the performance of DNFR and DSUR for both MU and MD time constraints and with respect to θ in Fig. 7. The obtained solutions are compared with the global minima for (CIA-UD) from pycombina. We observe that DNFR seems to perform better for MU time constraints, while DSUR performs better for the instances with MD time requirements. We plotted also the theoretical upper bound (UB) from Propositions 1 and 3, which are here 3 4 (C 1 +Δ), C 1 = C U , C D . As already observed for Fig. 6, the minima of (CIA-U) and (CIA-D) do hardly increase, if at all, for large MDTs and therefore diverge compared with the theoretical upper bound. We explain this behavior by the problem specific given relaxed values, which induce here an objective value of θ ≈ 0.87 for (CIA-U) and (CIA-D) even if no switches are used in the binary solution.
We show also the upper bound derived for DSUR with MU time constraints from Corollary 2, i.e. 5 6 (C U +Δ), and the lower bound for the upper bound for DSUR with MD time constraints by Proposition 4, i.e. 1 2 (C D +Δ). While the solutions constructed by DSUR may violate the upper bounds for (CIA-U) and (CIA-D), as happening for the MU time case, the bounds for DSUR are not violated. We observe that the (CIA-D) objective values are not only by far smaller than their upper bound, but even smaller or equal to the DSUR bound by Proposition 4.
We implemented DNFR and DSUR in Python 3.6 as additional solvers for pycombina. The execution of the heuristics took for all instances not more than 0.02 seconds on a workstation with 4 Intel i5-4210U CPUs (1.7 GHz) and 7.7 GB RAM. We conclude that the heuristics can be used to quickly generate robust solutions in terms of competitive objective values. They might also be useful for initializing pycombina's branch and bound algorithm with a good upper bound. However, a numerical study is needed for verifying the added benefit, which could be elaborated in future work.

Conclusions
In this article, we have derived tight bounds for the integrality gap of the CIA decomposition applied to MIOCPs under MDT constraints. The presented proofs are constructive and take advantage of the introduced DNFR scheme. Numerical experiments show that the CIA decomposition performs notably well in terms of objective quality for problems with small MDT requirements, since the deviation from its relaxed solution is negligible. For more restrictive dwell time constraints, the algorithm may provide feasible solutions that differ a lot from the relaxed solution. Hence, the constructed solution quality might be low compared to the optimal solution. Nevertheless, when considering the runtime of MINLP solvers, the CIA decomposition solution is computationally inexpensive and can be easily assessed with the relaxed solution for its quality. We have extended the SUR scheme so that it constructs dwell time feasible solutions and tested the proposed algorithm on a benchmark problem. The resulting (CIA) problem solutions of this and the DNFR scheme are close to the optimal ones. Hence, we propose that the DNFR or DSUR scheme may be beneficial in the context of huge (CIA) problems or in the setting of model predictive control, where a branch and bound algorithm struggles to find the optimal solution efficiently.
Acknowledgements Open Access funding provided by Projekt DEAL. The authors thank the anonymous reviewers for their helpful comments on earlier drafts of the manuscript that served to significantly improve and clarify the paper. This project has received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (Grant Agreement No 647573) and from Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)-314838170, GRK 2297 MathCoRe and SPP 1962.
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/.

A.1 Lemmata for DNFR Approximation Results
We present a series of lemmata, which will be needed later in the proofs to Theorems 1 and 2.
Proof Let us derive an auxiliary result for j ∈ [N ]: We use this and rearrange the sums in order to proof the first assertion: The auxiliary result is also useful for the second statement: Lemma 4 (Accumulated difference of Γ and Θover active controls) Let b 1 , b 2 ∈ [n b ] and we define S b 1 ,b 2 as the set of active controls between b 1 and b 2 : Then, we have Proof Using Definition 7 of Γ , Θ and rearranging sums yields Note that S b 1 ,b 2 is trivially the empty set, if b 2 ≤ b 1 + 1, but the result remains true in this case. We employ the concept of S b 1 ,b 2 for a contradiction in the proofs of Theorems 1 and 2.
Lemma 5 (Control with negative Γ value has not been future forced) Let (C 1 , C 2 , χ D ) be given and assume that the forward control deviation of a control i ∈ [n ω ] and a block b 2 ≥ 2 after executing Algorithm 1 satisfies: Then, there is an earlier activation of i on some block b 1 < b 2 and this activation has not been b 2 -future forced on b 1 .
Proof Note that Γ i,b is monotonically increasing in b for deactivated controls i. We conclude from this and Γ i,b 2 < 0 that there is an earlier activation of i on some block b 1 < b 2 . We take a closer look on the forward control deviation on block b 2 : and rearranging terms implies The last inequality shows us that i has been not b 2 -future forced on b 1 .

A.2 Proof of Theorem 1
Proof We proceed via induction. Base case: We consider the first b on which a down time forbidden control i D ∈ [n ω ] appears and let us assume holds and we prove that it results in a contradiction. It follows from Lemma 3 so that there must be a control i 1 = i D with negative forward control deviation on b: We apply Lemma 5 to the last inequality: i 1 has not been b-future forced on its last activation and we denote the block of this activation with b 1 . In other words, we know that there is at least one block b 1 and one control i 1 that was not b-future forced on b 1 and still was activated on b 1 . Now, we denote by i 1 the control of this property with the last activation before b. By this definition, we observe that all controls being activated after b 1 would become forced until b. We notate In particular, we have The last inequality holds, since control i was last activated at block k(i). For block b 1 we know that i 1 has been chosen, despite not being b-future forced. We use this observation and our assumption of b > b 1 being the first block with a down time forbidden control to conclude that all controls out of F b 1 We sum up the inequalities (A.2) and (A.3) over F b 1 ,b and similarly for (A.4), which yields (A.7) We used L b 1 , L b 2 ≤ L in the second inequality. We finish our calculations by considering the property of F b 1 ,b comprising all control activations between time blocks b 1 + 1 and b − 1, therefore we can apply Lemma 4 with F b 1 ,b = S b 1 ,b and obtain which is a contradiction to inequality (A.7).
Step case: Let the assertion hold until any block b − 1 ∈ [n b ] and we prove that then the statement holds for b. Again, we consider i D ∈ [n ω ] and assume holds and we prove that it results in a contradiction. With a similar argumentation as in the base case we deduce that there is a control i 1 that has not been b-future forced on block b 1 < b and reuse the definition of F b 1 ,b . Thus, inequality (A.3) still holds. Now, we distinguish between two cases why the controls out of F b 1 ,b have not been activated on b 1 . If all controls i ∈ F b 1 ,b have been inadmissible on b 1 , we can argue as in the base case. Hence, we focus on the other case: there is an i 2 ∈ F b 1 ,b , which was down time forbidden on b 1 and all other controls i ∈ F b 1 ,b \{i 2 } were inadmissible.
By the induction hypothesis and the previously derived inequality (A.4) we have Summing up these inequalities over F b 1 ,b results therefore in Next, we argue that F b 1 ,b contains at least two controls: the case b 1 = b − 1 is not possible, since i D is via assumption forced and down time forbidden on b, hence With this observation we subtract inequality (A.10) from (A.5): Notice that |F b 1 ,b | ≥ 2 is used in the last inequality. Finally, we build again on Lemma 4, where it is justified to set F b 1 ,b = S b 1 ,b . Thus, the above inequality is a contradiction to the inequality from the lemma and we have shown that the assertion holds for all b ∈ [n b ] on which a down time forbidden control exists.

A.3 Proof of Theorem 2
Proof The assertion can be shown for the parameter choices I. and II. in a very similar way, which is why we prove both cases in parallel. Since the algorithm activates for each block b ∈ [n ω ] either a forced, or future forced, or admissible control and the family of blocks is a partition [N ] by Lemma 2, exactly one control is activated on each interval j ∈ [N ] and therefore the (Conv) constraint satisfied. Hence, DNFR guarantees feasibility of w DNFR . If down time forbidden controls are neglected, i.e., χ D = 0, w DNFR yields an objective value with at most the claimed upper bound by the definition of admissible and forced activation. The same holds for the choice χ D = 1, since the control deviation does not become greater than the claimed upper bound during a MD time phase by Theorem 1. Therefore, we need only to prove that DNFR always provides a solution. For this, we show that for each interval there is (1.) at least one admissible control and (2.) at most one forced control.
(1.) We prove by contradiction that there exists at least one admissible control: Let us assume there is no admissible activation for block b ∈ [n b ] and distinguish between the cases: and we prove that it results in a contradiction. It follows from summing up all controls and from Lemma 3: and subtracting n ω L b from the right hand side yields II. If there is no down time forbidden control on b, we can proceed as in I. Otherwise, we may have one control i D that is down time forbidden. We assume all other controls are inadmissible, i.e., and we prove that it results in a contradiction. Hence, again by Lemma 3 and therefore The last inequality is a contradiction to Theorem 1.
We conclude that there must be an admissible activation for all blocks and thereby for all intervals. (2.) If there were more than one forced control at a time step, the algorithm would be ambiguous at line 3-4. Moreover, DNFR would provide, in this case, a solution that does not satisfy the upper bound on the objective. Therefore, we prove that this case is impossible and we do so again by contradiction. Assume that there b ∈ [n b ] is the block with smallest index on which at least two controls i 1 , i 2 are being forced, i.e., In the proof of Theorem 1 we have already shown how to obtain a contradiction with only one forward control deviation Γ i,b greater than the rounding threshold. So, case II. is settled with this theorem and we focus on case I. for which we proceed very similarly as in the proof of Theorem 1. Let us first apply Lemma 3: Hence, we have which implies that there is at least one control i 3 such that Then, by Lemma 5, there is an earlier activation of i 3 on some block b 3 < b and this activation has not been b-future forced on b 3 . Let i 3 denote the control of this property with the last activation before b. This definition implies that all controls that are active between b 3 and b become forced until b. We use again the notation In particular, we find i 1 , we apply the definition of F b 3 ,b and Γ : Since control i was last activated on block k(i) and b-future forced on k(i), we have For block b 3 we know that i 3 has been chosen, even though not being b-future forced. That implies i 3 was selected on b 3 because all controls out of F b 3 ,b were not admissible at this block. Hence, it results for i ∈ F b 3 ,b Now, we consider the sum of inequalities (A.11), (A.12) and sum up (A.13) over 15) from (A.14), we obtain In (A.16) we used L b 2 , L b 3 ≤ L , while the last inequality holds due to |F b 3 ,b | ≤ n ω −1. As in the proof of Theorem 1, we invoke now Lemma 4 with F b 3 ,b = S b 1 ,b in order to raise a contradiction to inequality (A.17). Overall, we have shown that there is at most one forced activation per block and thereby per interval. This completes the proof.

Remark 7
The proceeding in the proof of Theorem 2 shows us, on closer inspection, that DNFR provides a solution with control deviation bounded by C 2 L for the absence of MD time constraints, i.e., χ D = 0, and any chosen rounding threshold C 2 ≥ 2n ω −3 2n ω −2 and any block length parameter C 1 ≥ 0. This implies the previously known result by NFR [21], θ(w NFR ) ≤Δ, evolves as special case of DNFR with C 1 = 0, and C 2 = 1.

B Discussion on the Tightness of the obtained Minimum Down Time Bounds
Proposition 5 (Tightness of the bound for (CIA-D)) Let us assume the MD time constraint is active, i.e., C D > Δ is given. Then, the following is true: 1. The bound for (CIA-D) stated in Proposition 3 can not be improved by the DNFR scheme with χ D = 1 for n ω ≥ 3. 1 4 C D +Δ. Proof The assumption of an active MD time constraint ensures that the bound cannot be improved by the bound for MU times from Proposition 2. We use again the concept of a minimal C 1 -overlapping grid, here with C 1 = C D /2. 1. We want to prove that the DNFR scheme with χ D = 1 and C 2 < 3 2 may provide solutions with a (CIA-D) objective greater than C 2 L . Let us consider first C 2 ≤ 3 2 − ε 1 , with 0 < ε 1 ≤ 0.5. We present example values for a ∈ A with a time horizon length of at least 12C 1 , so that at least six blocks with length L exist by Lemma 1. Let 0 < ε 2 < ε 1 be small and let the relaxed control values a be given as

The bound for (CIA-D) is tight up to at most the constant
With this example values we discuss the thereby constructed DNFR solution as well its objective quality.
-First block: i 1 is 2-future forced and activated.
-Second block: Both i 1 and i 2 are 4-future forced. The DNFR algorithm breaks ties arbitrarily, so that activating i 2 is legitimate. -Third block: i 1 is down time forbidden, while i 2 is not admissible. DNFR activates therefore i 3 . -Fourth block: i 1 is activated, since it is forced. -Fifth block: We have in the meantime Θ i 1 ,4 = (0.5 + ε 1 − 2ε 2 )L and Θ i 2 ,4 = (0.5 − ε 1 + ε 2 )L . Since ε 2 satisfies ε 2 < ε 1 , both controls are 6-future forced on the fifth block. Let DNFR activate i 2 . -Sixth block: i 1 is still down time forbidden and can not be active, which implies so that the proposed control deviation bound is not fulfilled.
Finally, if ε 1 > 0.5, thus C 2 < 1, we can construct a similar example for which the control i 1 is already forced on the first block and the control deviation is again greater than C 2 L .
2. The MD time constraints are equivalent to MU time constraints with C U = C D for a problem with only two controls n ω = 2. Proposition 2 provides an example for this case, where θ ≥ 1 2 (C U +Δ) holds. This example can be also applied for more than two controls by setting the relaxed control values a i,b to zero, for i > 2. Then, the difference to the upper bound 3 4 C D + 3 2Δ from Proposition 3 is the one stated in the assertion.
Proposition 5 tells us the DNFR scheme with C 2 = 3 2 and χ D = 1 can not be improved. The following example motivates why we have chosen the set of the down time forbidden control in such a way that the active control can only be changed after a duration of C D /2 at the earliest. If it is possible to switch already after one interval Δ j , DNFR may construct greedy solutions with a large control deviation at long MD times C D . The following example shows the reason for this.

Example 1
Let n ω = 3 and a rounding threshold C 2 ≥ 2n ω −3 2n ω −2 be given. We alter DNFR in the following way: instead over blocks we iterate forward over all intervals. We keep for a given MD time C 1 = C D the threshold C 2 L for forced, future forced and admissible activation. In order to construct feasible solutions for (CIA-D), we extend the definition of I D b by letting all controls to be down time forbidden that are inactive and were active in the previous period of length C D . Next, we are going to construct exemplary relaxed values for this modified DNFR scheme with large control deviation. We first introduce recursively the indices where j 0 := 1. Let the relaxed values be given as follows 0 · · · 0 · · · 0 1 · · · 1 0 · · · 0 · · · 0 · · · 0 0 · · · 0 1 · · · 1 · · · ⎞ ⎟ ⎟ ⎠ .
Then, the modified DNFR may construct the following solution: where j 4 := min{ j ∈ [N ] | l∈[ j] Δ l < C D + Δ 1 } is the last index before the MD phase of i 1 ends. At first, i 1 is earliest future forced, but not anymore after being active on j = 1. Then, i 2 is activated on the second interval before i 3 is the earliest future forced control and needs to stay active until j 4 , since the other controls are down time forbidden. We notice that with small Δ 1 , Δ 2 it can result j 4 ≤ j 2 and therefore If we compare the term on the right inequality side with the bound from Proposition 3, i.e., 2n ω −3 2n ω −2 (C D +Δ) = 3 4 (C D +Δ), we conclude that the latter is less only if C D < 4Δ −Δ. Since Δ can be arbitrarily small and we assumed C D big compared with the grid length, the modified DNFR scheme would construct no improving bounds. Similar "greedy" examples can be constructed for n ω > 3 and block lengths greater than Δ.
Some comments on these tightness properties are in order.

Remark 8 (Quality of bound for (CIA-D))
The MD time configuration of DNFR, i.e., χ D = 1, yields smaller upper bounds compared with the DNFR algorithm with MU time configuration, i.e. χ D = 0 and C 1 = C D , only for instances with more than three controls and a large MD time C D compared with the grid lengthΔ. In fact, we conjecture the upper bound for any n ω to be θ max = 1 2 C D +Δ, therefore only slightly greater than the one for n ω = 2. With this threshold, there would be no forced control until the first down time forbidden control appears and we postulate that active controls that become forced without activation during the next C D time duration may stay active without other controls becoming forced. Of course, this argumentation justifies no proof -Proposition 5 together with Example 1 states that a generic solution fulfilling this bound can not be found via the DNFR scheme and it is presumably hard, if not even impossible, to construct it by another polynomial time algorithm.

Remark 9 (Quality of bound for (CIA-UD))
The integrality gap bound for (CIA-UD) as stated in Proposition 3 is tight for C U ≥ C D by the result of Proposition 2. For C U < C D , the bound is not necessarily tight, but it is again difficult to prove tight bounds due to the problem's combinatorial structure.

Remark 10
If we deal with an MDT C 1 that begins and ends exactly on the grid points, the upper bounds become 2n ω −3 2n ω −2 C U for (CIA-U), 3 4 C D for (CIA-D), and accordingly reduced for (CIA-UD).

C Proof of Proposition 4
Proof Using (6.1) and the definition of M D , we obtain M D ≥ 2(n ω − 1). Notice that even if C D /Δ / ∈ N, we still find for the cardinality of the dwell time index sets |J SUR k (C D )| = M D ∈ N for k ≤ N − M D because we deal with an equidistant grid. Hence, we calculate the forward control deviation of the currently activated control in the DSUR algorithm (line 3) on the next M D intervals.
k = 2: The control i = 1 is currently activated with a forward control deviation of Δ, calculated on the next M D intervals. The forward control deviation for control i = 2 amounts to γ 2,2 = θ 2,1 + a 2,2Δ = 0 +Δ. Therefore, DSUR may set the control i = 2 to be active. k = 3: We use the induction hypothesis for a (n ω −1) and Claim (C.1) that yields a (n ω −1) 2,M D +2 = 0. Thus, the forward control deviation of control i = 2 isΔ, which is the same for i = 3. We let DSUR deliberately set the control i = 3 to be active.
k = 4, . . . , n ω : We argue analogously to the case k = 3. We argued that this control deviation value is increased in w DSUR byΔ and before the choice w DSUR,(n ω −1) there exist M D + 1 columns in w DSUR . So, (C.3) is also true for n ω .