Minimizing the Number of Tardy Jobs and Maximal Tardiness on a Single Machine is NP-hard

Abstract

This paper resolves a long-standing open question in bicriteria scheduling regarding the complexity of a single machine scheduling problem which combines the number of tardy jobs and the maximal tardiness criteria. We use the lexicographic approach with the maximal tardiness being the primary criterion. Accordingly, the objective is to find, among all solutions minimizing the maximal tardiness, the one which has the minimum number of tardy jobs. The complexity of this problem has been open for over thirty years, and has been known since then to be one of the most challenging open questions in multicriteria scheduling. We resolve this question by proving that the problem is strongly NP-hard. We also prove that the problem is at least weakly NP-hard when we switch roles between the two criteria (i.e., when the number of tardy jobs is the primary criterion). Finally, we provide hardness results for two other approaches (constraint and a priori approaches) to deal with these two criteria.

1 Introduction

Since the early stages of classical scheduling theory, the main focus has been the optimization of a single criterion such as the makespan, maximal tardiness, total tardiness, or number of tardy jobs. However, in many practical cases, service and production organizations need to take more than a single objective into account when trying to produce an efficient schedule. For example, a production firm may want to balance its ability to meet job due dates with its ability to control the amount of work-in process held in the shop. Meeting the first objective may prioritize scheduling jobs with an early due date first, while meeting the second objective may prioritize scheduling jobs with short processing times first. As another example, consider a pizzeria that charges no money for late deliveries. Accordingly, the pizzeria owners may try to provide a delivery schedule that minimizes the number of tardy deliveries. However, such a strategy may yield an unfair solution when late deliveries have huge tardiness. In order to produce balanced delivery schedules, they may consider the maximal tardiness as an additional criterion to evaluate the quality of a delivery schedule. Given the above deficiency of traditional scheduling models, the field of multicriteria scheduling has gained a lot of attention from the late 80’s on. Ever since, the literature on multicriteria scheduling has expanded considerably, with several survey papers and books published over the years [2, 6, 10, 13, 14].

Consider two different minimization objectives $F_{1}$ and $F_{2}$ for a given scheduling problem. In a typical setting there are not enough resources to compute the entire set of Pareto optimal solutions, as this set is usually quite large. Therefore, one needs a way to define which solution is the most desired in this set. There are essentially three established approaches to tackle this issue, each of which defines a different problem for a given pair of scheduling criteria $F_{1}$ and $F_{2}$ (see, e.g., [6]).

• The lexicographic approach: Find a solution that minimizes $F_{2}$ among all solutions that minimize $F_{1}$. This variant is usually denoted by $1||Lex(F_{1},F_{2})$ in the single machine setting, where $F_{1}$ is called the primary criterion and $F_{2}$ is the secondary criterion.

• The constraint approach: Given a threshold parameter $\ell$, find a solution that minimizes $F_{1}$ subject to the constraint that $F_{2}\leq\ell$. In the single machine setting with no additional constraints this variant is usually denoted by $1|F_{2}\leq\ell|F_{1}$.

• The a priori approach: Find a solution that minimizes $\alpha F_{1}+F_{2}$, where $\alpha$ is a given constant that indicates the relative importance of criterion $F_{1}$ with respect to criterion $F_{2}$. This variant is usually denoted by $1||\alpha F_{1}+F_{2}$ in the single machine setting.

Note that the lexicographic approach is a special case of the a priori approach.

In this paper we consider two of the most basic and classical scheduling objectives: The first is the maximal tardiness criterion, typically denoted by $T_{\max}$, which measures the maximum tardiness of any job in the schedule. It is well-known that the $1||T_{\max}$ problem, the problem of minimizing $T_{\max}$ on a single machine, is solvable in $O(n\log n)$ time by processing the jobs based on the Earliest Due Date (EDD) rule [9]. The second criterion, denoted $\sum U_{j}$, is the total number of tardy jobs in the schedule. The corresponding single machine $1||\sum U_{j}$ problem is also solvable in $O(n\log n)$ time, due to an algorithm presented by Moore in the late 60s [11]. It follows that minimizing either the maximal tardiness or the number of tardy jobs can be done in polynomial time when the scheduling is done on a single machine. However, when both criteria are considered together, the resulting bicriteria problems become much harder to analyze and solve. In fact, the computational complexity status of the problem using either the constraint, lexicographic, or a priori approach was mentioned as an open problem by several different authors.

Lee and Vairaktarakis [10] published in 1993 an influential survey on bicriteria single machine scheduling problems. They focused on the lexicographic approach, and mentioned several open problems involving either one or both of the $T_{\max}$ and $\sum U_{j}$ criteria. Later on, all of these were resolved by Huo et al. [8, 7], apart from $1||Lex(T_{\max},\sum U_{j})$ and $1||Lex(\sum U_{j},T_{\max})$ which were left open. This is summarized in [7] with the following quote:

“Despite much efforts spent on $1||Lex(T_{\max},\sum U_{j})$ and $1||Lex(\sum U_{j},T_{\max})$, their complexity remain open. Although we cannot prove it, we conjecture that they are both NP-hard. It will be worthwhile to settle this issue in the future.”

The complexity status of the problem was mentioned as open also in later surveys by T’kindt and Billaut [13] and Hoogeveen [6], and also in the book on multicriteria scheduling by T’kindt and Billaut [14]. The $1|T_{\max}\leq\ell|\sum U_{j}$ problem is mentioned in the book on multiagent scheduling by Agnetis et al. [2], who write

“The complexity of this problem still stands out as one of the most prominent open issues in theoretical scheduling.”

As such, the complexity status of these problems have been established as one of the main open problems in multicriteria scheduling.

1.1 Our Results

We determine the computational complexity of single machine bicriteria scheduling involving objectives $T_{\max}$ and $\sum U_{j}$ using either the constraint, lexicographic, or a priori approach, by showing that all problems are unlikely to admit polynomial-time algorithms.

Our first main result involves the constraint variant of the problem. In its decision form, the problem asks to determine whether there exists a schedule with $T_{\max}\leq\ell$ and $\sum U_{j}\leq k$, and we denote this problem by $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$. We prove that this problem is strongly NP-complete by a reduction from 3-Partition.

Theorem 1.

$1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ is strongly NP-complete.

Following this, we show that there exists an easy reduction from $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ to the $1||Lex(T_{\max},\sum U_{j})$ problem, as well an easy reduction from $1||Lex(T_{\max},\sum U_{j})$ to the $1||\alpha T_{\max}+\sum U_{j}$ problem. As both reductions preserve strong NP-hardness, this directly yields:

Corollary 1.

$1||Lex(T_{\max},\sum U_{j})$ and $1||\alpha T_{\max}+\sum U_{j}$ are both strongly NP-hard.

Unfortunately, we could not find a direct reduction from $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ to $1||Lex(\sum U_{j},T_{\max})$. Thus, we are forced to design an alternative reduction for this last variant. While we cannot find such a reduction from a strong NP-hard problem, we are still able to devise a reduction from the weakly NP-complete Partition problem, giving us our second main result of the paper.

Theorem 2.

$1||Lex(\sum U_{j},T_{\max})$ is weakly NP-hard.

Thus, altogether our results resolve the complexity of all single machine bicriteria problems involving the $T_{\max}$ and $\sum U_{j}$ criteria, resolving the long standing open question posed in [2, 6, 7, 8, 10, 13, 14].

1.2 Related Work

Shantikumar [12] designed a branch-and-bound procedure for $1||Lex(\sum U_{j},T_{\max})$ which requires exponential time in the worst case. He also presented a polynomial-time algorithm for minimizing the maximal tardiness when the set of early jobs is given in advance. Chen and Bulfin [3] designed a branch-and-bound procedure for solving $1||Lex(T_{\max},\sum U_{j})$. By applying a set of numerical tests, they showed that their algorithm was able to solve instances of up to 40 jobs in less than one minute of computer time. Finally, Huo et al. [7] implemented several heuristics for $1||Lex(T_{\max},\sum U_{j})$ and tested them on instances with up to 200 jobs; their best heuristic was on average less than 1% worse than the optimal.

Huo et al. [7] also showed that the weighted problems $1||Lex(\max w_{j}T_{j},\sum U_{j})$ and $1||Lex(\sum U_{j},\max w_{j}T_{j})$, where each job has its own weight and one of the two criteria is to minimize the maximal weighted tardiness, are both weakly NP-hard. The results in this paper supersede both of these hardness results, as they are for the unweighted versions of these problems. The constraint problem $1|T_{\max}\leq\ell|\sum U_{j}$ is a special case of $1|\overline{d}_{j}|\sum U_{j}$. In this problem, each job has a due date $d_{j}$ as well as an additional deadline $\overline{d}_{j}$ which must be met. The goal is to minimize the number of tardy jobs while meeting all deadlines. When $\overline{d}_{j}=d_{j}+\ell$ for each job $j$, this problem becomes $1|T_{\max}\leq\ell|\sum U_{j}$. Yuan showed that $1|\overline{d}_{j}|\sum U_{j}$ is strongly NP-complete [15], and this result is superseded by our results as well.

2 Preliminaries

We use standard terminology from computer science and scheduling. For each $n\in\mathbb{N}$, we denote by $[n]:=\{1,\ldots,n\}$ the set of positive integers up to (and including) $n$. For a bijective function $f:X\rightarrow Y$, we denote its inverse by $f^{-1}$; thus, we have $f^{-1}(y)=x$ if and only if $f(x)=y$.

Scheduling.

We consider scheduling problems involving a set of $n$ jobs $\mathcal{J}$; where all are available at time zero to be non-preemptively processed on a single machine. Each job $J\in\mathcal{J}$ has two nonnegative parameters: its processing time $p(J)$ and its due date $d(J)$. A schedule is a bijection from $[n]$ to the set $\mathcal{J}$ of jobs, where the job $J\in\mathcal{J}$ with $\sigma(i)=J$ is the $i$-th job to be processed on the single machine.

Given a schedule $\sigma$, the completion time of job $J$ is $C_{\sigma}(J):=\sum_{i=1}^{\sigma^{-1}(J)}p(\sigma(i))$. We will focus on the following two objectives common in the scheduling literature: $T_{\sigma}(J):=\max\{0,C_{\sigma}(J)-d(J)\}$, which is the tardiness of job $J$; and $U_{\sigma}(J)$ which is a binary tardiness indicator that equals 1 if $C_{\sigma}(J)>d(J)$ and 0 otherwise. If the schedule $\sigma$ is clear from context, then we may also drop the subscript and just write $C(J)$, $T(J)$, or $U(J)$. If $U(J)=1$, then we say that job $J$ is tardy. Otherwise, $J$ is an early job.

We study the scheduling problems arising from combining the objective total number of tardy jobs ($\sum U_{j}$) with the maximum tardiness $T_{\max}=\max_{J\in\mathcal{J}}T(J)$ using the lexicographic, constraint, or a priori approach as described in Section 1. We will examplarily describe the problem resulting from the constrained approach using $T_{\max}$ as bounded objective:

$1|T_{\max}\leq\ell|\sum U_{j}$

Input:

A set $\mathcal{J}$ of jobs with processing times $p:\mathcal{J}\rightarrow\mathbb{N}$ and due dates $d:\mathcal{J}\rightarrow\mathbb{N}$ and an integer $\ell\in\mathbb{N}$.

Question:

Find a schedule $\sigma$ minimizing $\sum_{J\in\mathcal{J}}U(J)$ among all schedules with $\max_{J\in\mathcal{J}}T(J)\leq\ell$.

For the constrained version, we will call a schedule feasible if it obeys the constraint. For example, in the context of problem $1|T_{\max}\leq\ell|\sum U_{j}$, a schedule is feasible if $T(J)\leq\ell$ for every job $J\in\mathcal{J}$. A schedule is optimal if it is feasible and, among all feasible schedules, minimizes the objective. Taking again $1|T_{\max}\leq\ell|\sum U_{j}$ as an example, a schedule is optimal if $T(J)\leq\ell$ for every job $J\in\mathcal{J}$ and among all such schedules, it minimizes the number of tardy jobs. Note that the lexicographic version is a special case of the constrained version where the upper bound on the constrained objective is the minimum value of this objective over all schedules; thus, the notions of feasible and optimal schedule extend to the lexicographic case, where a schedule is feasible if it minimizes the primary objective.

Minimizing maximum tardiness for a given set of early jobs.

In order to minimize the maximum tardiness, one schedules the job according to non-decreasing due date [9]. This observation can easily be extended to finding a schedule with maximum tardiness $\ell$ subject to the condition that a given subset $\widetilde{\mathcal{J}}$ of jobs is early (see e.g. [5, Observation 3.2]). In order to do so, we defined a modified due date

$$\widetilde{d}(J):=\begin{cases}d(J)&\text{if }J\in\widetilde{\mathcal{J}}\\ d(J)+\ell&\text{if }J\notin\widetilde{\mathcal{J}}\end{cases}$$

Now scheduling all jobs according to non-decreasing $\widetilde{d}$ results in a schedule with a maximum tardiness of at most $\ell$ and each job from $\widetilde{\mathcal{J}}$ being early if such a schedule exists; we will call the corresponding schedule the canonical schedule for $\widetilde{\mathcal{J}}$.

Observation 1.

[5, Observation 3.2] Let $\mathcal{\widetilde{J}}$ be a subset of jobs. If there is a schedule with a total tardiness of at most $\ell$ where each job from $\widetilde{\mathcal{J}}$ is early, then the canonical schedule for $\widetilde{\mathcal{J}}$ is such a schedule.

Note that given a subset $\mathcal{\widetilde{J}}$ of jobs, a schedule minimizing $T_{\max}$ among all schedules where each job from $\widetilde{\mathcal{J}}$ is early can be found using Observation 1 and binary search.

3 Strong NP-hardness

This section describes a proof for Theorem 1: See 1 In order to show Theorem 1, we reduce from 3-Partition which is strongly NP-complete [4].

3-Partition

Input:

A multiset of $n$ integers $a_{1},\ldots,a_{n}$.

Question:

Is there a partition $(S_{1},\ldots,S_{m})$ of $[n]$ where $n=3m$ such that $\sum_{i\in S_{j}}a_{i}=t$ for every $j\in[m]$ where $t:=\sum_{i=1}^{n}a_{i}/m$?

We will interpret $1|\sum U_{j}\leq k,T_{\max}\leq\ell|$ as the decision version of the constrained $1|T_{\max}\leq\ell|\sum U_{j}$ problem, implying that the feasibility of a schedule refers to its maximum tardiness, i.e., we call a schedule feasible if its maximum tardiness is at most $\ell$. We will use throughout our construction a sufficiently large constant $\alpha$ defined by $\alpha=5n^{2}\cdot t$.

For ease of presentation, we partition the timeline into $m$ time periods, each corresponding to a different index $j\in[m]$. Define $\delta$ to be the value

$$\delta:=4n\cdot\alpha^{3}+2\cdot\alpha^{2}+(2m+1)t\cdot\alpha+mt\,.$$

The total length of each time period will be $\delta$, where the first period starts at time $\Delta_{0}=0$. In this way, the $j$’th time period starts at time $\Delta_{j-1}$ and ends at time $\Delta_{j}=\Delta_{j-1}+\delta$. Each period will consist of two halves. The length of the first half of the $j$’th time period will equal

$$\delta^{*}_{j}=2n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+1)t\cdot\alpha+(m-j)t\,.$$

In this way, the first half of the $j$’th time period ends at time $\Delta^{*}_{j}=\Delta_{j-1}+\delta^{*}_{j}$, which is the time when the second half starts. Note that the length of the first and second half of each period is not exactly but only roughly half of the length of the whole period (more specifically, each half has length half of the whole period if ignoring the lower-order terms multiplied by $\alpha$ or $\alpha^{0}$).

3.1 Construction

We next describe how to construct an instance of $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ from a given instance of 3-Partition. The high level idea is as follows: We will construct $m$ groups of jobs, one for each index $j\in[m]$, where the $j$’th group will encode all the integers in $a_{1},\ldots,a_{n}$ that are selected for the solution sets $S_{1}\cup\ldots\cup S_{j}$. Let $j\in[m]$. The $j$’th job group will consist of the $4n$ jobs $J_{i,j}$, $\neg J_{i,j}$, $J^{*}_{i,j}$, and $\neg J^{*}_{i,j}$ where $i\in[n]$. Intuitively speaking, scheduling $J_{i,j}$ and $J^{*}_{i,j}$ early will correspond to the situation where $a_{i}\in S_{1}\cup\ldots\cup S_{j}$, while scheduling $\neg J_{i,j}$ and $\neg J^{*}_{i,j}$ early corresponds to the situation where $a_{i}\notin S_{1}\cup\ldots\cup S_{j}$. The due date of the jobs are constructed in such a way that the due date of the last early job from $\{J_{n,j},\neg J_{n,j}\}$ will ensure that $\sum_{i\in S_{1}\cup\ldots\cup S_{j}}a_{i}\leq j\cdot t$, while the due date of the last early job from $\{J^{*}_{n,j},\neg J^{*}_{n,j}\}$ will ensure that $\sum_{i\in S_{1}\cup\ldots\cup S_{j}}a_{i}\geq j\cdot t$. Together, accounting for all $j\in[m]$, this will ensure that the sum of integers in each $S_{j}$ is exactly $t$.

Job construction.

For each $i\in[n]$ and each $j\in[m]$, we construct four number jobs $J_{i,j}$, $\neg J_{i,j}$, $J^{*}_{i,j}$, and $\neg J^{*}_{i,j}$, with the following processing times and due dates:

• $p(J^{*}_{i,j})=\alpha^{3}$ and $d(J^{*}_{i,j})=\Delta_{j-1}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$.

• $p(\neg J^{*}_{i,j})=\alpha^{3}+a_{i}$ and $d(\neg J_{i,j}^{*})=\Delta_{j-1}+(2i-1)\cdot\alpha^{3}+0.1\cdot\alpha^{2}$.

• $p(J_{i,j})=\alpha^{3}+a_{i}\cdot\alpha$ and $d(J_{i,j})=\Delta^{*}_{j}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$.

• $p(\neg J_{i,j})=\alpha^{3}$ and $d(\neg J_{i,j})=\Delta^{*}_{j}+(2i-1)\cdot\alpha^{3}+0.1\cdot\alpha^{2}$.

Note that we have

$$d(J_{n,j-1}) <d(\neg J_{1,j}^{*})<d(J_{1,j}^{*})<d(\neg J_{2,j}^{*})<d(J_{2,j}^{*})<\ldots<d(\neg J_{n,j}^{*})<d(J_{n,j}^{*})$$

$$<d(\neg J_{1,j})<d(J_{1,j})<d(\neg J_{2,j})<d(J_{2,j})<\ldots<d(\neg J_{n,j})<d(J_{n,j})<d(\neg J_{1,j+1}^{*}).$$

Furthermore, for each $j\in[m]$, we add two delimiter jobs $D_{j}^{*}$ and $D_{j}$ in addition to the $4n$ jobs defined above. These jobs are used to indicate the end of the first and second halves of the $j$-th period. That is, time $\Delta^{*}_{j}$ and time $\Delta_{j}$, respectively. The characteristics of $D_{j}^{*}$ and $D_{j}$ are defined as follows:

• $p(D_{j}^{*})=\alpha^{2}+(m-j)t\cdot\alpha$ and $d(D_{j}^{*})=\Delta^{*}_{j}$, and

• $p(D_{j})=\alpha^{2}+jt\cdot\alpha$ and $d(D_{j})=\Delta_{j}$.

We will also need to add some filler jobs for the first and last time period. We start by constructing the job $F_{0}$ with processing time and due date $p(F_{0})=d(F_{0})=m\cdot t\cdot\alpha$. This job will always be scheduled first in any feasible schedule for the entire $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ instance. Furthermore, we construct $n$ filler jobs $F^{1}_{1},\ldots,F^{1}_{n}$ for the first time period, and $n$ filler jobs $F^{m}_{1},\ldots,F^{m}_{n}$ for the last time period. These filler jobs will have the following characteristics:

• $p(F_{i}^{1})=\alpha^{3}$ and due date $d(F_{i}^{1})=d(J^{*}_{i,1})$ for each $i\in[n]$ (so $F_{i}^{1}$ is a copy of $J^{*}_{i,1}$), and

• $p(F_{i}^{m})=\alpha^{3}$ and $d(F_{i}^{m})=\Delta_{m}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$ for each $i\in[n]$.

Job	Processing Time	Due Date
$J^{*}_{i,j}$	$\alpha^{3}$	$\Delta_{j-1}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$
$\neg J^{*}_{i,j}$	$\alpha^{3}+a_{i}$	$\Delta_{j-1}+(2i-1)\cdot\alpha^{3}+0.1\cdot\alpha^{2}$
$D_{j}^{*}$	$\quad\alpha^{2}+(m-j)t\cdot\alpha$	$\Delta^{*}_{j}$
$J_{i,j}$	$\alpha^{3}+a_{i}\cdot\alpha$	$\Delta^{*}_{j}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$
$\neg J_{i,j}$	$\alpha^{3}$	$\Delta^{*}_{j}+(2i-1)\cdot\alpha^{3}+0.1\cdot\alpha^{2}$
$D_{j}$	$\alpha^{2}+jt\cdot\alpha$	$\Delta_{j}$
$F_{0}$	$mt\cdot\alpha$	$mt\cdot\alpha$
$F_{i}^{1}$	$\alpha^{3}$	$d(J^{*}_{i,1})$
$F_{i}^{m}$	$\alpha^{3}$	$\Delta_{m}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}$

Table 1: The processing times and due dates of all jobs constructed for the $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ instance.

Parameters $\boldmath{k}$ and $\boldmath{\ell}$.

We have described above all the jobs constructed for the $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ instance. For an overview of their processing times and due dates, see Table 1. To finish our construction, we set the target number $k$ of tardy jobs to

$$k\,\,=\,\,2m\cdot n,$$

and the target maximum tardiness $\ell$ to

$$\ell\,\,=\,\,2n\cdot\alpha^{3}+\alpha^{2}+0.1\cdot\alpha^{2}.$$

Intuitively, the target maximum tardiness allows each job to be tardy by approximately half a period.

3.2 Job sets $\boldmath{\mathcal{J}_{\leq j}}$ and $\boldmath{\mathcal{J}^{*}_{\leq j}}$

To prove the correctness of our reduction we will make frequent use of the following sets of jobs: For each $j\in[m]$, we use $\mathcal{J}^{*}_{j}:=\{J_{i,j}^{*},\neg J_{i,j}^{*}:i\in[n]\}$ and $\mathcal{J}_{j}:=\{J_{i,j},\neg J_{i,j}:i\in[n]\}$. Next, we define $\mathcal{J}^{*}_{0}=\emptyset$ and $\mathcal{J}_{0}=\{F_{0}\}\cup\{F_{i}^{1}:i\in[n]\}$. Then, for each $j\in[m]$, we define

$$\mathcal{J}^{*}_{\leq j}:=\mathcal{J}_{\leq j-1}\cup\mathcal{J}^{*}_{j}\cup\{D^{*}_{j}\}$$

and

$$\mathcal{J}_{\leq j}:=\mathcal{J}_{\leq j}^{*}\cup\mathcal{J}_{j}\cup\{D_{j}\}.$$

Note that for $j>0$, $\mathcal{J}_{\leq j}$ contains all jobs with due dates at most $\Delta_{j}$, whereas $\mathcal{J}_{\leq j}^{*}$ contains all jobs with due dates at most $\Delta_{j}^{*}$.

We first observe that all jobs from $\mathcal{J}_{\leq j}^{*}$ except $D_{j}^{*}$ must be completed by the end of the $j$-th period, i.e., by time $\Delta_{j}$, as

Observation 2.

$$\max_{J\in\mathcal{J}_{\leq j}^{*}\setminus\{D_{j}^{*}\}}d(J)+\ell<\Delta_{j}.$$

Proof.

The observation follows by easy calculations:

$$\max_{J\in\mathcal{J}_{\leq j}^{*}\setminus\{D_{j}^{*}\}}d(J)+\ell =d(J^{*}_{n,j})+\ell$$

$$=\Delta_{j-1}+2n\cdot\alpha^{3}+0.1\cdot\alpha^{2}+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j-1}+4n\cdot\alpha^{3}+1.2\cdot\alpha^{2}$$

$$=\Delta_{j}-0.8\cdot\alpha^{2}-(2m+1)t\cdot\alpha-mt<\Delta_{j}.\qed$$ (1)

Similarly, we observe that all jobs from $\mathcal{J}_{\leq j}$ except for $D_{j}$ must be completed by $\Delta_{j+1}^{*}$ as

Observation 3.

$$\max_{J\in\mathcal{J}_{\leq j}\setminus\{D_{j}\}}d(J)+\ell<\Delta_{j+1}^{*}.$$

Proof.

The proof follows by calculations analogous to the proof of Observation 2:

$$\max_{J\in\mathcal{J}_{\leq j}\setminus\{D_{j}\}}d(J)+\ell =d(J_{n,j})+\ell$$

$$=\Delta_{j}^{*}+2n\cdot\alpha^{3}+0.1\cdot\alpha^{2}+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j}^{*}+4n\cdot\alpha^{3}+1.2\cdot\alpha^{2}$$

$$=\Delta_{j+1}^{*}-0.8\cdot\alpha^{2}-(2m-1)t\cdot\alpha-(m-1)t<\Delta_{j+1}^{*}.\qed$$

We next compare the total processing time of all jobs in $\mathcal{J}_{\leq j}^{*}$ to $\Delta_{j}$. This value represents, assuming that all jobs from $\mathcal{J}_{\leq j}^{*}$ are completed by the end of the $j$-th period (which is true in all feasible schedules for all jobs from $\mathcal{J}_{\leq j}^{*}$ except for $D_{j}^{*}$ by Observation 2), how much processing time is left for other jobs. The resulting identity will be used frequently:

Observation 4.

For each $j\in[m]$, we have

$$\Delta_{j}=p(\mathcal{J}^{*}_{\leq j})+n\cdot\alpha^{3}+\alpha^{2}+2jt\cdot\alpha\,.$$

Proof.

The proof is by induction on $j$. For $j=1$, we have

$$p(\mathcal{J}_{\leq 1}^{*}) =\sum_{i=1}^{n}\bigl{(}p(J^{*}_{i,1})+p(\neg J^{*}_{i,1})\bigr{)}+p(D^{*}_{1})+\sum_{i=1}^{n}p(F_{i}^{1})+p(F_{0})$$

$$=2n\cdot\alpha^{3}+mt+\alpha^{2}+(m-1)t\cdot\alpha+n\cdot\alpha^{3}+mt\cdot\alpha$$

$$=3n\cdot\alpha^{3}+\alpha^{2}+(2m-1)t\cdot\alpha+m\cdot t=\Delta_{1}-n\cdot\alpha^{3}-\alpha^{2}-2t\cdot\alpha\,.$$

For $j>1$, note that

$$\mathcal{J}_{\leq j}^{*}\setminus\mathcal{J}^{*}_{\leq j-1}=\mathcal{J}^{*}_{j}\cup\mathcal{J}_{j-1}\cup\{D_{j}^{*},D_{j-1}\}\,.$$

Thus, we have (using induction for the second equality)

$$p(\mathcal{J}^{*}_{\leq j}) =p(\mathcal{J}^{*}_{\leq j-1})+p(\mathcal{J}^{*}_{\leq j}\setminus\mathcal{J}^{*}_{\leq j-1})$$

$$=\Delta_{j-1}-\Bigr{(}n\cdot\alpha^{3}+\alpha^{2}+2(j-1)t\cdot\alpha\Bigr{)}+p(\mathcal{J}_{j}^{*})+p(\mathcal{J}_{j-1})+p(D_{j}^{*})+p(D_{j-1})$$

$$=\Delta_{j-1}-n\cdot\alpha^{3}-\alpha^{2}-2(j-1)t\cdot\alpha+2n\cdot\alpha^{3}+mt+2n\cdot\alpha^{3}+mt\cdot\alpha$$

$$\qquad+\alpha^{2}+(m-j)t\cdot\alpha+\alpha^{2}+(j-1)t\cdot\alpha$$

$$=\Delta_{j-1}+3n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+1)t\cdot\alpha+mt$$

$$=\Delta_{j}-n\cdot\alpha^{3}-\alpha^{2}-2jt\cdot\alpha\,.\qed$$

Similarly, it will be useful to compute the difference between the end of the first half of the $j$’th period and total processing time of $\mathcal{J}_{\leq j-1}$:

Observation 5.

For each $j\in[n]$, we have

$$\Delta^{*}_{j}=p(\mathcal{J}_{\leq j-1})+n\cdot\alpha^{3}+\alpha^{2}+(m-j)t\cdot\alpha+(m-j)t\,.$$

Proof.

We have (using Observation 4 for the second equality)

$$p(\mathcal{J}_{\leq j-1}) =p(\mathcal{J}_{\leq j}^{*})-\Bigl{(}p(\mathcal{J}_{j}^{*})+p(D_{j}^{*})\Bigr{)}$$

$$=\Delta_{j}-\Bigl{(}n\cdot\alpha^{3}+\alpha^{2}+2jt\cdot\alpha\Bigr{)}-\Bigl{(}2n\cdot\alpha^{3}+mt+\alpha^{2}+(m-j)t\cdot\alpha\Bigr{)}$$

$$=\Delta_{j}-\Bigl{(}3n\cdot\alpha^{3}+2\cdot\alpha^{2}+(m+j)t\cdot\alpha+mt\Bigr{)}$$

$$=\Delta_{j-1}+\delta-\Bigl{(}3n\cdot\alpha^{3}+2\cdot\alpha^{2}+(m+j)t\cdot\alpha+mt\Bigr{)}$$

$$=\Delta_{j}^{*}+2n\cdot\alpha^{3}+\alpha^{2}+2jt\cdot\alpha+jt-\Bigl{(}3n\cdot\alpha^{3}+2\cdot\alpha^{2}+(m+j)t\cdot\alpha+mt\Bigr{)}$$

$$=\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)\cdot t\,.\qed$$

3.3 Candidate Sets

We will now consider a special kind of subsets of jobs which we will call candidate sets. As we will later see in Section 3.4, it suffices to consider canonical schedules for candidates sets. We start by defining a candidate set of jobs:

Definition 1.

A candidate set is a set $\widetilde{\mathcal{J}}$ of jobs such that for each $i\in[n]$ and $j\in[m]$, exactly one of $J^{*}_{i,j}$ and $\neg J^{*}_{i,j}$, and exactly one of $J_{i,j}$ and $\neg J_{i,j}$ is contained in $\widetilde{\mathcal{J}}$, and every filler or delimiter job is contained in $\widetilde{\mathcal{J}}$.

Whenever we will talk in the following about the canonical schedule for a candidate set $\widetilde{\mathcal{J}}$, the possible tie between $F_{i}^{1}$ and $J_{i,1}^{*}$ will always be broken in favor of $J_{i,1}^{*}$ i.e., if both $F_{i}^{1}$ and $J_{i,1}^{*}$ are contained in $\widetilde{\mathcal{J}}$, then $F_{i}^{1}$ will be scheduled after $J_{i,1}^{*}$ in the canonical schedule for $\widetilde{\mathcal{J}}$. We remark that not all candidate sets lead to feasible schedules.

Example 1.

Assume that $n=4$, $m=2$, and

$$\widetilde{\mathcal{J}}=\mathcal{J}_{F}\cup\mathcal{J}_{D}\cup\{J^{*}_{1,1},\neg J^{*}_{2,1},\neg J^{*}_{3,1},J^{*}_{4,1},J^{*}_{1,2},J^{*}_{2,2},J^{*}_{3,2},J^{*}_{4,2}\}\cup\{J_{1,1},\neg J_{2,1},\neg J_{3,1},J_{4,1},J_{1,2},J_{2,2},J_{3,2},J_{4,2}\}$$

where $\mathcal{J}_{F}$ is the set of filler jobs and $\mathcal{J}_{D}$ is the set of delimiter jobs. Then the candidate schedule for $\widetilde{\mathcal{J}}$ is

$$\begin{array}[]{rrrrrrrrrrrrrrrrrrrr}F_{0},&J^{*}_{1,1},&F_{1}^{1},&\neg J^{*}_{2,1},&F_{2}^{1},&\neg J^{*}_{3,1},&F_{3}^{1},&J^{*}_{4,1},&F_{4}^{1},&D^{*}_{1}\\ &\neg J^{*}_{1,1},&J_{1,1},&\neg J_{2,1},&J^{*}_{2,1},&\neg J_{3,1},&J^{*}_{3,1},&\neg J^{*}_{4,1},&J_{4,1},&D_{1}\\ \hline\cr&\neg J_{1,1},&J^{*}_{1,2},&J^{*}_{2,2},&J_{2,1},&J^{*}_{3,2},&J_{3,1},&\neg J_{4,1},&J^{*}_{4,2},&D^{*}_{2}\\ &\neg J^{*}_{1,2},&J_{1,2},&\neg J^{*}_{2,2},&J_{2,2},&\neg J^{*}_{3,2},&J_{3,2},&\neg J^{*}_{4,2},&J_{4,2},&D_{2}\\ \hline\cr&\neg J_{1,2},&F_{1}^{m},&\neg J_{2,2},&F_{2}^{m},&\neg J_{3,2},&F_{3}^{m},&\neg J_{4,2},&F_{4}^{m}\,.\end{array}$$

The horizontal lines separate the time periods, and each row (except for the last one) represent one half of a period.

We continue by analyzing canonical schedules for candidate sets. More precisely, we want to classify the early jobs in a canonical schedule for a candidate set and determine when the canonical schedule for a candidate set is feasible. We start characterizing which candidate sets result in a feasible schedule.

Lemma 1.

Let $\widetilde{\mathcal{J}}$ be a candidate set. Then the canonical schedule for $\widetilde{\mathcal{J}}$ is feasible if and only if there is no $i\in[n]$ and $j\in[m-1]$ such that $\neg J^{*}_{i,j+1},J_{i,j}\in\widetilde{\mathcal{J}}$ and there is no $i\in[n]$ and $j\in[m]$ such that $\neg J_{i,j},J^{*}_{i,j}\in\widetilde{\mathcal{J}}$.

Proof.

First note that jobs $F_{0}$ and $F_{i}^{1}$ are completed by time $\ell$ (independent on the set $\widetilde{\mathcal{J}}$) and thus have tardiness smaller than $\ell$.

Next we will show that job $J^{*}_{i,j}$ for $i\in[n]$ and $j\in[m]$ always has tardiness at most $\ell$. We assume, without loss of generality, that $J^{*}_{i,j}\notin\widetilde{\mathcal{J}}$ (otherwise $J^{*}_{i,j}$ will be completed earlier than in the following calculation). Before $J^{*}_{i,j}$, all jobs from $\mathcal{J}_{\leq j-1}$ as well as $n+i$ jobs from $\mathcal{J}^{*}_{j}$ (including $J^{*}_{i,j}$, and each with processing time at most $\alpha^{3}+t$), $D^{*}_{j}$, and $i$ jobs (each with processing time at most $p(J)\leq\alpha^{3}+t\cdot\alpha$) from $\mathcal{J}_{j}$ are scheduled. Thus, using Observation 5, $J^{*}_{i,j}$ is completed by

$$p(\mathcal{J}_{\leq j-1})+ (n+i)\cdot(\alpha^{3}+t)+\alpha^{2}+(m-j)t\cdot\alpha+i\cdot(\alpha^{3}+t\cdot\alpha)$$

$$=\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)t+(n+2i)\cdot\alpha^{3}+\alpha^{2}+(i+m-j)t\cdot\alpha+(n+i)t$$

$$=\Delta_{j}^{*}+2i\cdot\alpha^{3}+it\cdot\alpha+(n+i+j-m)t$$

$$=\Delta_{j-1}+2n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+1)t\cdot\alpha+(m-j)\cdot t+2i\cdot\alpha^{3}+it\cdot\alpha+(n+i+j-m)t$$

$$<\Delta_{j-1}+2(n+i)\cdot\alpha^{3}+\alpha^{2}+0.1\cdot\alpha^{2}=d(J^{*}_{i,j})+2n\cdot\alpha^{3}+\alpha^{2}<d(J^{*}_{i,j})+\ell\,.$$

We continue by characterizing when job $\neg J^{*}_{i,j}$ for $i\in[n]$ and $j\in[m]$ has tardiness at most $\ell$. We assume, without loss of generality, that $\neg J^{*}_{i,j}\notin\widetilde{\mathcal{J}}$ (otherwise $\neg J^{*}_{i,j}$ will be completed much earlier than in the following calculations). Before $\neg J^{*}_{i,j}$, all jobs from $\mathcal{J}_{\leq j-1}$ as well as $n+i$ jobs from $\mathcal{J}^{*}_{j}$ (including $\neg J^{*}_{i,j}$ itself), $D_{j}^{*}$, and $i$ or $i-1$ jobs from $\mathcal{J}_{j}$ (depending on whether $J_{i,j}\in\widetilde{\mathcal{J}}$) are scheduled. We continue by showing that (under our assumption $\neg J^{*}_{i,j}\notin\widetilde{\mathcal{J}}$) job $\neg J^{*}_{i,j}$ has tardiness at most $\ell$ if and only if $J_{i,j}\in\widetilde{\mathcal{J}}$. If $J_{i,j}\in\widetilde{\mathcal{J}}$, then there are only $i-1$ jobs from $\mathcal{J}_{j}$ before $\neg J^{*}_{i,j}$ and consequently, using the analogous calculations as for $J^{*}_{i,j}$ (note that the only difference to $J^{*}_{i,j}$ is that there are now $i-1$ instead of $i$ jobs from $\mathcal{J}_{j}$), job $\neg J^{*}_{i,j}$ is completed by time

$$\Delta_{j-1}+(2n+2i-1)\cdot\alpha^{3}+\alpha^{2}+0.1\cdot\alpha^{2}=d(\neg J^{*}_{i,j})+2n\cdot\alpha^{3}+\alpha^{2}<d(\neg J^{*}_{i,j})+\ell.$$

If, however, $\neg J_{i,j}\in\widetilde{\mathcal{J}}$, then there are $i$ jobs from $\mathcal{J}_{j}$ before $\neg J^{*}_{i,j}$ and thus $\neg J^{*}_{i,j}$ is not completed before (using the same calculations as for $J^{*}_{i,j}$)

$$p(\mathcal{J}_{\leq j-1})+ (n+i)\cdot\alpha^{3}+\alpha^{2}+(m-j)t\cdot\alpha+i\cdot\alpha^{3}$$

$$=\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)t+(n+2i)\cdot\alpha^{3}+\alpha^{2}+(m-j)t\cdot\alpha$$

$$=\Delta_{j}^{*}+2i\cdot\alpha^{3}-(m-j)t$$

$$=\Delta_{j-1}+2n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+1)t\cdot\alpha+(m-j)t+2i\cdot\alpha^{3}-(m-j)t$$

$$=d(\neg J^{*}_{i,j})+(2n+1)\cdot\alpha^{3}+0.9\cdot\alpha^{2}+(2m-2j+1)t\cdot\alpha$$

$$>d(\neg J^{*}_{i,j})+\ell\,.$$

Consequently, $\neg J^{*}_{i,j}$ has tardiness more than $\ell$ if and only if $J_{i,j}^{*}\in\widetilde{\mathcal{J}}$ and $\neg J_{i,j}\in\widetilde{\mathcal{J}}$.

The arguments for $J_{i,j}$ and $\neg J_{i,j}$ are analogous: Before $J_{i,j}$ (assume, without loss of generality, that $\neg J_{i,j}\in\widetilde{\mathcal{J}}$), all jobs from $\mathcal{J}_{\leq j}^{*}$ as well as $n+i$ jobs from $\mathcal{J}_{j}$ (including $J_{i,j}$), $D_{j}$, and $i$ jobs from $\mathcal{J}^{*}_{j+1}$ are scheduled. Thus, using Observation 4, $J_{i,j}$ is completed by time

$$\Delta_{j}^{*}+2(n+i)\cdot\alpha^{3}+\alpha^{2}+0.1\cdot\alpha^{2}<d(J_{i,j})+\ell.$$

For job $\neg J_{i,j}$, again assume, without loss of generality, that $J_{i,j}\in\widetilde{\mathcal{J}}$. First note that there are $i-1$ jobs from $\mathcal{J}^{*}_{j+1}$ scheduled before $\neg J_{i,j}$ if $J^{*}_{i,j+1}\in\widetilde{\mathcal{J}}$, and $i$ jobs from $\mathcal{J}^{*}_{j+1}$ otherwise. Analogous calculations to the case for $\neg J_{i,j}$ now show that $\neg J_{i,j}$ has tardiness at most $\ell$ if and only if $J_{i,j}\in\widetilde{\mathcal{J}}$ and $\neg J^{*}_{i,j+1}\in\widetilde{\mathcal{J}}$.

We continue with job $D_{j}^{*}$. This job is scheduled after $\mathcal{J}_{\leq j-1}$ and $n$ jobs from $\mathcal{J}^{*}_{j}$. Thus, using Observation 5, it is completed by time

$$\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)t+n\cdot(\alpha^{3}+t)+\alpha^{2}+(m-j)t\cdot\alpha=\Delta_{j}^{*}+(n+j-m)\cdot t<d(D_{j}^{*})+\ell.$$

Similarly, $D_{j}$ is scheduled after $\mathcal{J}_{\leq j}^{*}$ and $n$ jobs from $\mathcal{J}_{j}$ and thus (using Observation 4) completed by time

$$\Delta_{j}-n\cdot\alpha^{3}-\alpha^{2}-2jt\cdot\alpha+n\cdot(\alpha^{3}+t\cdot\alpha)+\alpha^{2}+jt\cdot\alpha=\Delta_{j}+(n-j)t\cdot\alpha<d(D_{j})+\ell.\qed$$

Next, let us identify which jobs are early in the canonical schedule for a candidate set. We start with the filler jobs, which turn out to be early for every candidate set:

Lemma 2.

In the canonical schedule for any candidate set, jobs $F_{0}$, $F_{i}^{1}$, and $F_{i}^{m}$ for $i\in[n]$ are early.

Proof.

Job $F_{0}$ is the first job and thus completed at time $p(F_{0})=d(F_{0})$. Until job $F_{i}^{1}$, job $F_{0}$, jobs $F_{i_{0}}^{1}$ for $i_{0}\leq i$, and $i$ number jobs are scheduled. Thus, $F_{i}^{1}$ is completed by time

$$p(F_{0})+i\cdot p(F_{i}^{1})+i\cdot\max_{J}p(J)<2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}=d(F_{i}^{1})\,.$$

Finally, let us consider job $F_{i}^{m}$. This job is completed after $\mathcal{J}^{*}_{\leq m}$, jobs $J_{i_{0},m}$ and $\neg J_{i_{0},m}$ for $i_{0}\leq i$, exactly one of $J_{i_{0},m}$ and $\neg J_{i_{0},m}$ for $i_{0}>m$, job $D_{m}$, and $F_{i_{0}}^{m}$ for $i_{0}\leq i$. Thus, using Observation 4, $F_{i}^{m}$ is completed by time

$$p(\mathcal{J}^{*}_{\leq m}) +\sum_{i_{0}=1}^{i}\bigl{(}p(J_{i_{0},m})+p(\neg J_{i_{0},m})\bigr{)}+\sum_{i_{0}=i+1}^{n}p(J_{i_{0},m})+p(D_{m})+\sum_{i_{0}=1}^{i}p(F_{i_{0}}^{m})$$

$$\leq\Delta_{m}-n\cdot\alpha^{3}-\alpha^{2}-2mt\cdot\alpha+(n+i)\cdot\alpha^{3}+mt\cdot\alpha+\alpha^{2}+mt\cdot\alpha+i\cdot\alpha^{3}$$

$$=\Delta_{m}+2i\cdot\alpha^{3}<d(F_{i}^{m})\,.\qed$$

We continue with characterizing when number jobs are early.

Lemma 3.

In the canonical schedule for a candidate set $\widetilde{\mathcal{J}}$, a number job $J$ is early if and only if $J\in\widetilde{\mathcal{J}}$.

Proof.

Easy calculations similar to Lemma 1 show the statement: We start with $J^{*}_{i,j}$, and first assume that $J^{*}_{i,j}\in\widetilde{\mathcal{J}}$. Job $J^{*}_{i,j}$ is scheduled after $\mathcal{J}^{*}_{\leq j-1}$, $n+i-1$ or $n+i$ jobs from $\mathcal{J}_{j-1}$ (depending on whether $J_{i,j-1}\in\widetilde{\mathcal{J}}$), $D_{j-1}$, and $i$ jobs from $\mathcal{J}^{*}_{j}$ (including $J^{*}_{i,j}$). Consequently, using Observation 4, $J^{*}_{i,j}$ is completed by time

$$p(\mathcal{J}^{*}_{\leq j-1})+ (n+i)\cdot(\alpha^{3}+t\cdot\alpha)+\alpha^{2}+(j-1)t\cdot\alpha+i\cdot(\alpha^{3}+t)$$

$$=\Delta_{j-1}-n\cdot\alpha^{3}-\alpha^{2}-2(j-1)t\cdot\alpha+(n+2i)\cdot\alpha^{3}+\alpha^{2}+(n+i+j-1)t\cdot\alpha+it$$

$$=\Delta_{j-1}+2i\cdot\alpha^{3}+(n+i-j+1)t\cdot\alpha+it<d(J^{*}_{i,j})\,.$$

If $J^{*}_{i,j}\notin\widetilde{\mathcal{J}}$, then $J^{*}_{i,j}$ is scheduled after all jobs from $\mathcal{J}_{\leq j-1}$ as well as $n+i$ jobs from $\mathcal{J}^{*}_{j}$ (including $J^{*}_{i,j}$, and each with processing time at least $\alpha^{3}$), $D^{*}_{j}$, and $i$ jobs (each with processing time at least $p(J)\geq\alpha^{3}$) from $\mathcal{J}_{j}$. Consequently, using Observation 5, $J^{*}_{i,j}$ is completed not before time

$$p(\mathcal{J}_{\leq j-1})+ (n+i)\cdot\alpha^{3}+\alpha^{2}+(m-j)t\cdot\alpha+i\cdot\alpha^{3}$$

$$=\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)t+(n+2i)\cdot\alpha^{3}+\alpha^{2}+(m-j)t\cdot\alpha$$

$$=\Delta_{j}^{*}+2i\cdot\alpha^{3}-(m+j)t$$

$$=\Delta_{j-1}+2n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+1)t\cdot\alpha+(m-j)\cdot t+2i\cdot\alpha^{3}+(m-j)t$$

$$>\Delta_{j-1}+2i\cdot\alpha^{3}+0.1\cdot\alpha^{2}=d(J^{*}_{i,j})\,.$$

The calculations for $\neg J^{*}_{i,j}$, $J_{i,j}$ and $\neg J_{i,j}$ are analogous (using Observation 5 instead of Observation 4 for $J_{i,j}$ and $\neg J_{i,j}$). ∎

We continue with identifying when a delimiter job is early in a candidate schedule:

Lemma 4.

Let $\sigma$ be the canonical schedule for some candidate set $\widetilde{\mathcal{J}}$. Then $D_{j}^{*}$ is early if and only if $\sum_{i:J^{*}_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\geq jt$. Further, $D_{j}$ is early if and only if $\sum_{i:J_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\leq jt$.

Proof.

Up to $D_{j}^{*}$, the following jobs are scheduled: $\mathcal{J}_{\leq j-1}$, $J^{*}_{i,j}$ if $J^{*}_{i,j}\in\widetilde{\mathcal{J}}$, $\neg J^{*}_{i,j}$ if $\neg J^{*}_{i,j}\in\widetilde{\mathcal{J}}$, and $D_{j}^{*}$. Using Observation 5, the total processing time of these jobs is

$$p(\mathcal{J}_{\leq j-1}) +\sum_{i:J^{*}_{i,j}\in\widetilde{\mathcal{J}}}p(J^{*}_{i,j})+\sum_{i:\neg J^{*}_{i,j}\in\widetilde{\mathcal{J}}}p(\neg J^{*}_{i,j})+p(D_{j}^{*})$$

$$=\Delta_{j}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j)t\cdot\alpha-(m-j)t+n\cdot\alpha^{3}+\sum_{i:\neg J^{*}_{i,j}\in\widetilde{\mathcal{J}}}a_{i}+\alpha^{2}+(m-j)t\cdot\alpha$$

$$=d(D_{j}^{*})+\sum_{i:\neg J^{*}_{i,j}\in\widetilde{\mathcal{J}}}a_{i}-(m-j)t\,.$$

Consequently, $D_{j}^{*}$ is early if and only if $\sum_{i:\neg J^{*}_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\leq(m-j)t$. Since $\sum_{i=1}^{n}a_{i}=mt$, this is equivalent to $\sum_{i:J^{*}_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\geq jt$.

We now turn to the second part of the lemma. Up to $D_{j}$, the following jobs are scheduled: $\mathcal{J}^{*}_{\leq j}$, $J_{i,j}$ if $J_{i,j}\in\widetilde{\mathcal{J}}$, $\neg J_{i,j}$ if $\neg J_{i,j}\in\widetilde{\mathcal{J}}$, and $D_{j}$. Using Observation 4, the total processing time of these jobs is

$$p(\mathcal{J}^{*}_{\leq j}) +\sum_{i:J_{i,j}\in\widetilde{\mathcal{J}}}p(J_{i,j})+\sum_{i:\neg J_{i,j}\in\widetilde{\mathcal{J}}}p(\neg J_{i,j})+p(D_{j})$$

$$=\Delta_{j}-n\cdot\alpha^{3}-\alpha^{2}-2jt\cdot\alpha+n\cdot\alpha^{3}+\sum_{i:J_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\cdot\alpha+\alpha^{2}+jt\cdot\alpha$$

$$=d(D_{j})+\sum_{i:J_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\cdot\alpha-jt\cdot\alpha\,.$$

Consequently, $D_{j}$ is early if and only if $\sum_{i:J_{i,j}\in\widetilde{\mathcal{J}}}a_{i}\leq jt$. ∎

3.4 Correctness

In order to show the correctness of the reduction described in this section, the first (and hardest) part is to show that we can restrict ourselves to considering candidate schedules. This will be done in the next lemma:

Lemma 5.

If there is a feasible schedule with $k$ tardy jobs, then there is an optimal schedule that is the canonical schedule for a candidate set.

Before proving Lemma 5, we first show how to derive the main result in this section, namely the strong NP-hardness of $1|\sum U_{j}\leq k,T_{\max}\leq\ell|$:

Proof of Theorem 1.

$1|T_{\max}\leq\ell,U_{j}\leq k|$ is clearly contained in NP as an optimal schedule is a certificate. It remains to show that the problem is NP-hard. We will do so via the reduction from 3-Partition as described throughout this section. The reduction clearly runs in polynomial time, so it remains to show its correctness.

We start with the forward direction. So let $(S_{1},\ldots,S_{m})$ be a solution to the 3-Partition instance. Let

$$\widetilde{\mathcal{J}}:=\{J^{*}_{i,j},J_{i,j}:a_{i}\in S_{1}\cup\ldots\cup S_{j}\}\cup\{F_{0},F_{i}^{1},F_{i}^{m}:i\in[n]\}\cup\{D_{j},D_{j}^{*}:j\in[m]\},$$

and let $\sigma$ be the canonical schedule for $\widetilde{\mathcal{J}}$. By Lemma 1, $\sigma$ is feasible. By Lemma 2, jobs $F_{0}$, $F_{i}^{1}$, and $F_{i}^{m}$ for $i\in[n]$ are early. By Lemma 3, we have $2nm$ early and $2nm$ tardy number jobs. By Lemma 4 and since we have a solution to 3-Partition, all early delimiter jobs are early. Overall, we have $k=2mn$ tardy jobs, finishing the forward direction.

We continue with the backward direction. So let $\sigma$ be a feasible schedule with at most $k$ tardy jobs. By Lemma 5, we may assume that $\sigma$ is the canonical schedule for some candidate set $\widetilde{\mathcal{J}}$. By Lemma 3, there are $2nm$ tardy number jobs. This implies that $2nm$ number jobs and all non-number jobs are early. Let $I_{j}^{*}:=\{i\in[n]:J^{*}_{i,j}\in\widetilde{\mathcal{J}}\}$ and $I_{j}:=\{i\in[n]:J_{i,j}\in\widetilde{\mathcal{J}}\}$. Since $\sigma$ is feasible, Lemma 1 implies that whenever $J_{i,j}^{*}\in\mathcal{\widetilde{J}}$, then also $J_{i,j}\in\mathcal{\widetilde{J}}$ (as otherwise $\neg J_{i,j},J_{i,j}^{*}\in\widetilde{\mathcal{J}}$, contradicting the feasibility of $\sigma$ by Lemma 1). In other words, we have $I_{j}^{*}\subseteq I_{j}$. Similarly, the feasibility of $\sigma$ together with Lemma 1 implies that whenever $J_{i,j}\in\widetilde{\mathcal{J}}$, then also $J_{i,j+1}^{*}\in\widetilde{\mathcal{J}}$, i.e., $I_{j}\subseteq I_{j+1}^{*}$. Because every delimiter job is early, Lemma 4 implies that $\sum_{i\in I_{j}^{*}}a_{i}\geq jt$ and $\sum_{i\in I_{j}}a_{i}\leq jt$. Combining the above two results, we have

$$\sum_{i\in I_{j}^{*}}a_{i}\geq jt\geq\sum_{i\in I_{j}}a_{i}\,.$$

Since $I_{j}^{*}\subseteq I_{j}$, this implies that $I_{j}^{*}=I_{j}$ and that the above inequality holds with equality. We claim that $(S_{1},\ldots,S_{m}):=(I_{1}^{*},I_{2}^{*}\setminus I_{1}^{*},\ldots,I_{m}^{*}\setminus I_{m-1}^{*})$ is a solution to the 3-Partition instance. First note that $S_{j_{0}}$ and $S_{j_{1}}$ are disjoint for $j_{0}\neq j_{1}$ as $I_{1}^{*}=I_{1}\subseteq I_{2}^{*}=I_{2}\subseteq I_{3}^{*}=I_{3}\subseteq\ldots\subseteq I_{m}^{*}\subseteq I_{m}$. Further, we have

$$\sum_{i\in S_{j}}a_{i}=\sum_{i\in I_{j}^{*}}a_{i}-\sum_{i\in I_{j-1}^{*}}a_{i}=jt-(j-1)t=t\,.\qed$$

The rest of this section focuses on the proof of Lemma 5, i.e., that there always is an optimal canonical schedule for some candidate set $\widetilde{\mathcal{J}}$ if an optimal schedule has $k$ early jobs. In order to show Lemma 5, we start with very coarse statements on an optimal schedule and then refine them step by step. Using Observation 1, we may assume that there is a canonical optimal schedule. We observe a very coarse structure of any canonical feasible schedule:

Lemma 6.

Any canonical feasible schedule $\sigma$ satisfies that

• for that every $j\in[m]$, jobs from $\mathcal{J}^{*}_{j}$ and $D_{j}^{*}$ are scheduled after every job from $\mathcal{J}^{*}_{\leq j-1}$ and before any job not contained in $\mathcal{J}_{\leq j}$, and

• for every $j\geq 2$, jobs from $\mathcal{J}_{j}$ and $D_{j}$ are scheduled after every job from $\mathcal{J}_{\leq j-1}$ and before every job not contained in $\mathcal{J}^{*}_{\leq j+1}$.

Proof.

We start with the first bullet point. Note that the maximal due date in $\mathcal{J}^{*}_{\leq j-1}$ belongs to $D^{*}_{j-1}$. Thus, for each job $J\in\mathcal{J}^{*}_{\leq j-1}$, we have

$$\widetilde{d}_{\sigma}(J) \leq d(D^{*}_{j-1})+\ell$$

$$=\Delta_{j-1}^{*}+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j-2}+\delta_{j-1}^{*}+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j-2}+2n\cdot\alpha^{3}+\alpha^{2}+(2m-2j+3)t\cdot\alpha+(m-j+1)t+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j-2}+4n\cdot\alpha^{3}+2.1\cdot\alpha^{2}+(2m-2j+3)t\cdot\alpha+(m-j+1)t$$

$$=\Delta_{j-1}+0.1\cdot\alpha^{2}+(2-2j)t\cdot\alpha+(1-j)\cdot t<d(\neg J^{*}_{i,j})<d(J^{*}_{i,j})<d(D_{j}^{*})\,.$$

Furthermore, for any $J\notin\mathcal{J}_{\leq j}$, we have

$$\max\{\widetilde{d}_{\sigma}(\neg J^{*}_{i,j}),\widetilde{d}_{\sigma}(J^{*}_{i,j}),\widetilde{d}_{\sigma}(D_{j}^{*})\}\leq d(D^{*}_{j})+\ell=\Delta_{j}^{*}+\ell=\Delta_{j}+0.1\cdot\alpha^{2}-2jt\cdot\alpha-jt<d(J)\,.$$

The first bullet point now follows as $\sigma$ is canonical.

The proof of the second bullet point is similar to the one of the first bullet point. For each $J\in\mathcal{J}_{\leq j-1}$, we have

$$\widetilde{d}_{\sigma}(J) \leq d(D_{j-1})+\ell$$

$$=\Delta_{j-1}+2n\cdot\alpha^{3}+1.1\cdot\alpha^{2}$$

$$=\Delta_{j-1}+\delta_{j}^{*}+0.1\cdot\alpha^{2}-(2m-2j+1)\cdot\alpha-(m-j)t$$

$$=\Delta_{j}^{*}+0.1\cdot\alpha^{2}-(2m-2j+1)\cdot\alpha-(m-j)t<d(\neg J_{i,j})<d(J_{i,j})<d(D_{j})\,.$$

Furthermore, for any $J\notin\mathcal{J}^{*}_{\leq j+1}$, we have

$$\max\{\widetilde{d}_{\sigma}(\neg J_{i,j}),\widetilde{d}_{\sigma}(J_{i,j}),\widetilde{d}_{\sigma}(D_{j})\}\leq d(D_{j})+\ell=\Delta_{j}+\ell=\Delta_{j+1}^{*}+0.1\cdot\alpha^{2}-(2m-2j-1)t\cdot\alpha-(m-j-1)t<d(J)\,.$$

The second bullet point now follows as $\sigma$ is canonical. ∎

In the following, we mostly focus on the jobs with processing time at least $\alpha^{3}$, i.e., all jobs but the delimiter jobs and $F_{0}$. Thus, we will call a job large if its processing time is at least $\alpha^{3}$. All other jobs are called small. Therefore, the set of large jobs is $\{J^{*}_{i,j},\neg J^{*}_{i,j},J_{i,j},\neg J_{i,j}:i\in[n],j\in[m]\}\cup\{F_{i}^{1},F_{i}^{m}:i\in[n]\}$. As a consequence of Lemma 6, there are strong restrictions on which large job may be scheduled in which time period:

Lemma 7.

Any canonical feasible schedule $\sigma$ satisfies that the only large jobs which can be scheduled in

• the first half of the first period are $F_{i}^{1}$ for $i\in[n]$ and $\mathcal{J}^{*}_{1}\cup\mathcal{J}_{1}$,

• the first half of the $j$-th period for $j>1$ are $\mathcal{J}^{*}_{j}\cup\mathcal{J}_{j-1}$, and

• the second half of the $j$-th period are $\mathcal{J}^{*}_{j}\cup\mathcal{J}_{j}$.

Proof.

We start with the first bullet point. By Lemma 6, for $j\geq 2$, jobs $\mathcal{J}^{*}_{j}\cup\mathcal{J}_{j}$ are all scheduled after all jobs from $\mathcal{J}^{*}_{\leq 1}$. Since $p(\mathcal{J}^{*}_{\leq 1})=3n\cdot\alpha^{3}+\alpha^{2}+(2m-1)t\cdot\alpha+mt>\Delta_{1}^{*}$, this implies that all these jobs are scheduled after the first half of the first period.

We continue with the second bullet point. Let $j\geq 2$. Applying Lemma 6 for $j_{0}>j$, $\mathcal{J}^{*}_{j_{0}}\cup\mathcal{J}_{j_{0}}$ are all scheduled after all jobs from $\mathcal{J}^{*}_{\leq j}$. Since $p(\mathcal{J}^{*}_{\leq j})=\Delta_{j}-n\cdot\alpha^{3}-\alpha^{2}-2jt\cdot\alpha>\Delta_{j}^{*}$ by Observation 4, this implies that all jobs from $\mathcal{J}^{*}_{j_{0}}\cup\mathcal{J}_{j_{0}}$ for $j_{0}>j$ are scheduled after the first half of the $j$’th period. Further, any large job $J$ from $\mathcal{J}^{*}_{\leq j-1}$ must be completed by the end of the $(j-1)$’th period by Observation 2. Consequently, the only remaining large jobs which can be scheduled in the first half of the $j$’th period are $J^{*}_{i,j}$, $\neg J^{*}_{i,j}$, $J_{i,j-1}$, and $\neg J_{i,j-1}$.

Lastly, we show the third bullet point analogously to the second bullet point. Applying Lemma 6, jobs $\mathcal{J}^{*}_{j_{0}}$ for $j_{0}>j+1$ and $\mathcal{J}_{j_{0}}$ for $j_{0}>j$ are all scheduled after all jobs from $\mathcal{J}_{\leq j}$. Since $p(\mathcal{J}_{\leq j})=\Delta_{j+1}^{*}-n\cdot\alpha^{3}-\alpha^{2}-(m-j-1)t\cdot\alpha-(m-j-1)t>\Delta_{j}$ by Observation 5, this implies that all jobs from $\mathcal{J}^{*}_{j_{0}}$ for $j_{0}>j+1$ or $\mathcal{J}_{j_{0}}$ for $j_{0}>j$ are scheduled after the $j$’th period. Further, any large job $J$ from $\mathcal{J}_{\leq j-1}$ must be completed by the end of the first half of the $j$’th period by Observation 3. Consequently, the only remaining large jobs which can be scheduled in the second half of the $j$’th period are $J^{*}_{i,j}$, $\neg J^{*}_{i,j}$, $J_{i,j}$, and $\neg J_{i,j}$. ∎

We now bound the number of early jobs from each set $\mathcal{J}^{*}_{j}$ or $\mathcal{J}_{j}$.

Lemma 8.

For any feasible schedule,

• up to $2n$ jobs from $\{F_{i}^{1},J^{*}_{i,1},\neg J^{*}_{i,1}:i\in[n]\}$ can be completed by $\Delta_{1}^{*}$,

• up to $n$ jobs from $\mathcal{J}^{*}_{j}$ can be completed by $\Delta_{j}^{*}$ for any $j\geq 2$, and

• up to $n$ jobs from $\mathcal{J}_{j}$ can be completed by $\Delta_{j}$ for $j\in[m]$.

Proof.

Let $\sigma$ be a canonical feasible schedule. We first show the first bullet point. Note that each job from $\{F_{i}^{1},J^{*}_{i,1},\neg J^{*}_{i,1}:i\in[n]\}$ is large, i.e., has processing time at least $\alpha^{3}$. Thus, at most $2n$ jobs from $\{F_{i}^{1},J^{*}_{i,1},\neg J^{*}_{i,1}:i\in[n]\}$ can be completed by $\Delta_{1}^{*}<(2n+1)\cdot\alpha^{3}$.

We continue with the second bullet point. By Observation 3, all jobs from $\mathcal{J}_{\leq j-1}\setminus\{D_{j-1}\}$ are completed by $\Delta_{j}^{*}$. Thus, the total processing time of all jobs from $\mathcal{J}^{*}_{j}$ that are completed by $\Delta_{j}^{*}$ is upper-bounded by $\Delta_{j}^{*}-p(\mathcal{J}_{\leq j-1}\setminus\{D_{j-1}\})=n\cdot\alpha^{3}+2\cdot\alpha^{2}+(m-1)t\cdot\alpha+(m-j)t<(n+1)\cdot\alpha^{3}$ using Observation 5 for the equality. Because each job from $\mathcal{J}^{*}_{j}$ has a processing time of at least $\alpha^{3}$, this implies that at most $n$ jobs from $\mathcal{J}^{*}_{j}$ can be completed by $\Delta_{j}^{*}$.

We conclude with the third bullet point which is analogous to the second bullet point. By Observation 2, all jobs from $\mathcal{J}^{*}_{\leq j}\setminus\{D^{*}_{j}\}$ are completed by time $\Delta_{j}$. Thus, the total processing time of all jobs from $\mathcal{J}_{j}$ which are completed by time $\Delta_{j}$ is upper-bounded by $\Delta_{j}-p(\mathcal{J}^{*}_{\leq j}\setminus\{D^{*}_{j}\})=n\cdot\alpha^{3}+2\cdot\alpha^{2}+(m+j)t\cdot\alpha<(n+1)\cdot\alpha^{3}$ using Observation 4 for the equality. Because each job from $\mathcal{J}_{j}$ has processing time at least $\alpha^{3}$, this implies that at most $n$ jobs from $\mathcal{J}_{j}$ can be completed by time $\Delta_{j}$. ∎

As an easy consequence of Lemma 8, we get restrictions on the set of early jobs in an optimal solution:

Lemma 9.

For any feasible schedule with at most $k$ tardy jobs, the set of tardy jobs looks as follows:

• $2n$ jobs from $\{F_{i}^{1}:i\in[n]\}\cup\mathcal{J}^{*}_{1}$ are tardy, and these tardy jobs are completed after $\Delta_{1}^{*}$,

• $n$ jobs from $\mathcal{J}^{*}_{j}$ are tardy for any $j\geq 2$, and these tardy jobs are completed after $\Delta_{j}^{*}$, and

• $n$ jobs from $\mathcal{J}_{j}$ are tardy for $j\in[m]$, and all these tardy jobs are completed after $\Delta_{j}$.

Proof.

Note that each job from $\mathcal{J}^{*}_{j}$ that is early must be completed by time $\Delta_{j}^{*}$. Further, each job from $\mathcal{J}_{j}$ that is early must be completed by time $\Delta_{j}$. If $F_{i}^{1}$ is early, then it is completed by time $d(F_{i}^{1})<\Delta_{1}^{*}$. Thus, Lemma 8 implies that for each $j\geq 2$, there are $2n$ tardy jobs from $\mathcal{J}^{*}_{j}\cup\mathcal{J}_{j}$, and there are $2n$ tardy jobs from $\mathcal{J}^{*}_{1}\cup\mathcal{J}_{1}\cup\{F_{i}^{1}:i\in[n]\}$. Since there are at most $k=2mn$ tardy jobs overall, it follows that for each $\mathcal{J}^{*}_{j}$ or $\mathcal{J}_{j}$ with the exception of $\mathcal{J}^{*}_{1}$, there are exactly $n$ tardy jobs.

The second part of each bullet point follows from Lemma 8 which imply that at most $n$ jobs from $\mathcal{J}^{*}_{j}$ respectively $\mathcal{J}_{j}$ can be scheduled up to time $\Delta_{j}^{*}$ respectively $\Delta_{j}$. ∎

We can finally show Lemma 5:

Proof of Lemma 5.

Let $\sigma$ be an optimal schedule. By Observation 1, we may assume that $\sigma$ is the canonical schedule for some set $\widetilde{\mathcal{J}}$. Since $F_{i}^{1}$ and $J^{*}_{i,1}$ have the same characteristics, we may assume, without loss of generality, that if at least one of $F_{i}^{1}$ and $J^{*}_{i,1}$ is early, then $F_{i}^{1}$ is early. It remains to show that $\widetilde{\mathcal{J}}$ is a candidate set, i.e., that $\mathcal{\widetilde{J}}$ contains every filler job, every delimiter job, and for each $i\in[n]$ and $j\in[m]$, exactly one of $J_{i,j}^{*}$ and $\neg J_{i,j}^{*}$ and exactly one of $J_{i,j}$ and $\neg J_{i,j}$. Lemma 9 together with the assumption that if one of $F_{i}^{1}$ and $J_{i,1}$ is early, then $F_{i}^{1}$ is early implies that $\widetilde{\mathcal{J}}$ contains every filler and every delimiter job.

It remains to show that $\widetilde{\mathcal{J}}$ contains for each $i\in[n]$ and $j\in[m]$, exactly one of $J_{i,j}^{*}$ and $\neg J_{i,j}^{*}$ and exactly one of $J_{i,j}$ and $\neg J_{i,j}$. We prove this by backwards induction on $i$. Let $i=n$, and assume towards a contradiction that $\widetilde{\mathcal{J}}$ contains both $J_{i,j}$ and $\neg J_{i,j}$ or both $J_{i,j}^{*}$ and $\neg J_{i,j}^{*}$ (we will handle the other case that $\widetilde{\mathcal{J}}$ contains neither $J_{i,j}$ nor $\neg J_{i,j}$ or neither $J_{i,j}^{*}$ nor $\neg J_{i,j}^{*}$ later). If both $J^{*}_{n,j},\neg J^{*}_{n,j}\in\widetilde{\mathcal{J}}$ for some $j\in[m]$, then also $J_{n,j}$ and $\neg J_{n,j}$ need to be early and thus contained in $\widetilde{\mathcal{J}}$: By Observations 2 and 9 and since $d(J)<\Delta_{j}$ for every $J\in\mathcal{J}_{j}$, the jobs from $\mathcal{J}^{*}_{\leq j}$ and $n$ early jobs from $\mathcal{J}_{j}$ are scheduled by time $\Delta_{j}$. For all jobs $J\in\mathcal{J}^{*}_{\leq j}\setminus\{J^{*}_{n,j},\neg J^{*}_{n,j}\}$, it holds by Observation 2 that

$$d(J)+\ell\leq d(J^{*}_{n-1,j})+\ell=d(J^{*}_{n,j})-2\cdot\alpha^{3}+\ell<\Delta_{j}-2\cdot\alpha^{3}\,.$$

Consequently, no job from $\mathcal{J}^{*}_{\leq j}$ is completed after $\Delta_{j}-2\cdot\alpha^{3}$ (note that $J^{*}_{n,j}$ and $\neg J^{*}_{n,j}$ are early and thus completed by $d(J^{*}_{n,j})<\Delta_{j}-2\cdot\alpha^{3}$). The only two jobs from $\mathcal{J}_{j}$ with due date at least $\Delta_{j}-2\cdot\alpha^{3}$ are $J_{n,j}$ and $\neg J_{n,j}$. Consequently, $J_{n,j}$ and $\neg J_{n,j}$ are the only candidates from $\mathcal{J}_{j}$ to be completed between time $\Delta_{j}-2\cdot\alpha^{3}$ and $\Delta_{j}$. Thus, they are completed between $\Delta_{j}-2\cdot\alpha^{3}$ and $\Delta_{j}$, and they are early as any job from $\mathcal{J}_{j}$ that is completed by $\Delta_{j}$ is early by Lemma 9.

Analogous arguments show that if both $J_{n,j}$ and $\neg J_{n,j}$ are contained in $\widetilde{\mathcal{J}}$, then both $J^{*}_{n,j+1}$ and $\neg J^{*}_{n,j+1}$ are also contained in $\widetilde{\mathcal{J}}$.

However, it cannot be the case that both $J_{n,m}$ and $\neg J_{n,m}$ are contained in $\widetilde{\mathcal{J}}$: As each non-number job is early by Lemma 9, $F_{i_{0}}^{m}$ is early for each $i_{0}\in[n]$. Consequently, the only jobs which may be completed after time $\Delta_{m}+(2n-2)\cdot\alpha^{3}+0.1\cdot\alpha^{2}$ are $F_{n}^{m}$ and jobs $J_{n,m}$ or $\neg J_{n,m}$ if they are tardy. As at least two jobs must be scheduled after $\Delta_{m}+(2n-2)\cdot\alpha^{3}$ (because the total processing time of all jobs together is $\Delta_{m}+2n\cdot\alpha^{3}$), this implies that at least one of $J_{n,m}$ or $\neg J_{n,m}$ is tardy.

Now assume that neither $J^{*}_{n,j}$ nor $\neg J^{*}_{n,j}$ or neither $J_{n,j}$ nor $\neg J_{n,j}$ are contained in $\widetilde{\mathcal{J}}$. We will arrive at a contradiction in three steps, similar to the case that both $J^{*}_{n,j}$ and $\neg J^{*}_{n,j}$ are contained in $\widetilde{\mathcal{J}}$. First, we show that $J^{*}_{n,j_{0}}$ and $\neg J^{*}_{n,j_{0}}$ being tardy implies that $J_{n,j_{0}-1}$ and $\neg J_{n,j_{0}-1}$ are tardy for every $j_{0}\in\{2,\ldots,m\}$. Second, we show that $J_{n,j_{0}}$ and $\neg J_{n,j_{0}}$ being tardy implies that $J^{*}_{n,j_{0}}$ and $\neg J^{*}_{n,j_{0}}$ are tardy for every $j_{0}\in[m]$. Third, we will show that both $J^{*}_{n,1}$ and $\neg J^{*}_{n,1}$ being tardy leads to a contradiction. We start with the first step by showing that the only large jobs which may be completed between time $\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ and $\Delta_{j_{0}}^{*}$ are $J_{n,j_{0}-1}$ and $\neg J_{n,j_{0}-1}$. As $d(J)<\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ for every $J\in\mathcal{J}^{*}_{j_{0}}\setminus\{J^{*}_{n,{j_{0}}},\neg J^{*}_{n,{j_{0}}}\}$, each job from $\mathcal{J}^{*}_{j_{0}}$ that is completed after $\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ is tardy (note that $J^{*}_{n,j_{0}}$ and $\neg J^{*}_{n,j_{0}}$ are tardy by assumption). By Lemma 9, the tardy jobs from $\mathcal{J}^{*}_{j_{0}}$ are completed after $\Delta_{j_{0}}^{*}$. By Lemma 7, this implies that only jobs from $\mathcal{J}_{j_{0}-1}$ can be scheduled between $\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ and $\Delta_{j_{0}}^{*}$. The only two jobs $J$ from $\mathcal{J}_{j_{0}-1}$ with $d(J)+\ell\geq\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ are $J_{n,j_{0}-1}$ and $\neg J_{n,j_{0}-1}$, implying that these two jobs are completed between $\Delta_{j_{0}}^{*}-2\cdot\alpha^{3}$ and $\Delta_{j_{0}}^{*}$ and therefore are tardy. We continue with the second step. Similar arguments show that $J_{n,j_{0}}$ and $\neg J_{n,j_{0}}$ being tardy implies that $J^{*}_{n,j_{0}}$ and $\neg J^{*}_{n,j_{0}}$ are tardy. We finally do the third step: Repeatedly applying the previous two steps implies $J^{*}_{n,1}$ and $\neg J^{*}_{n,1}$ are tardy. By Lemma 9, there are $2n$ early jobs from $\{J^{*}_{i,1},\neg J^{*}_{i,1},F_{i}^{1}:i\in[n]\}$. Of these jobs, only $J^{*}_{n,1}$, $\neg J^{*}_{n,1}$, and $F_{i}^{1}$ have a due date larger than $\Delta_{1}^{*}-2\cdot\alpha^{3}$. Consequently, two of these three jobs are early. Since we assume that whenever $J^{*}_{i,1}$ is early, $F_{i}^{1}$ is also early which implies that $F_{n}^{1}$ and exactly one of $J^{*}_{n,1}$ and $\neg J^{*}_{n,1}$ are early.

The induction step is analogous, observing that directly before $D_{j}^{*}$, exactly one of $J^{*}_{i_{0},j}$ and $\neg J^{*}_{i_{0},j}$ and exactly one of $J_{i_{0},j-1}$ and $\neg J_{i_{0},j-1}$ for $i_{0}>i$ are scheduled and directly before $D_{j}$, exactly one of $J^{*}_{i_{0},j}$ and $\neg J^{*}_{i_{0},j}$ and exactly one of $J_{i_{0},j}$ and $\neg J_{i_{0},j}$ for $i_{0}>i$ are scheduled. ∎

4 Lexicographically first minimizing $T_{\max}$ and then $\sum U_{j}$

In this section, we show that $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ can be reduced to the problem of first minimizing the maximum tardiness and then the number of tardy jobs:

Theorem 3.

$1||Lex(T_{\max},\sum U_{j})$ is strongly NP-hard.

Proof.

We reduce from $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ which is strongly NP-hard by Theorem 1. Let $\mathcal{I}=(\mathcal{J}=\{J_{1},\ldots,J_{n}\},\ell,k)$ be an instance of $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$. Let $P:=\sum_{i=1}^{n}p(J_{i})$. We assume, without loss of generality, that $\ell<P$ and $d(J)\leq P$ for every $J\in\mathcal{J}$, and that there is some schedule with maximum tardiness at most $\ell$. We create a job $J^{*}$ with $p(J^{*})=P$ and $d(J^{*})=2P-\ell$. We claim that $\mathcal{I}^{\prime}:=(\mathcal{J}\cup\{J^{*}\},k+1)$ is a yes-instance of $1||Lex(T_{\max},\sum U_{j})$ if and only if $\mathcal{I}$ is a yes-instance of $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$.

As a first step of showing the correctness of the reduction, we show that the minimal maximum tardiness for $\mathcal{I}^{\prime}$ is precisely $\ell$. Let $\sigma^{\prime}$ be any schedule for $\mathcal{I}^{\prime}$. If $J^{*}$ is not the last job in $\sigma^{\prime}$, then the last job $J$ in $\sigma^{\prime}$ is completed at time $2P$ and thus has tardiness at least $2P-d(J)\geq P>\ell$. Otherwise $J^{*}$ is the last job in $\sigma^{\prime}$ and has tardiness exactly $2P-d(J^{*})=\ell$, implying that every schedule for $\mathcal{I}^{\prime}$ has maximum tardiness at least $\ell$. To show that the minimum maximum tardiness is exactly $\ell$, it suffices to describe a schedule with maximum tardiness $\ell$. Note that taking the presumed schedule for $\mathcal{I}$ with maximum tardiness at most $\ell$ and appending $J^{*}$ is such a schedule.

We conclude the prove by showing correctness. We start with the forward direction. Given a solution $\sigma$ to $\mathcal{I}$, appending $J^{*}$ to $\sigma$ results in a schedule for $\mathcal{I}^{\prime}$ with one more tardy job (namely $J^{*}$) and maximum tardiness $\ell$.

Given a solution $\sigma^{\prime}$ to $\mathcal{I}^{\prime}$, first note that any schedule minimizing $T_{\max}$ must schedule job $J^{*}$ last. Consequently, job $J^{*}$ is tardy. Further, the minimal maximum tardiness of a schedule is exactly $\ell$, so job $J$ has tardiness at most $\ell$ for every $J\in\mathcal{J}$. Thus, $\sigma\setminus\{J^{*}\}$ has maximum tardiness at most $\ell$ and at most $k$ tardy jobs. ∎

We will now show that the a priori version is strongly NP-hard:

Corollary 2.

For any $\alpha>0$, $1||T_{\max}+\alpha\sum U_{j}$ is strongly NP-hard.

Proof.

We reduce from $1||Lex(T_{\max},\sum U_{j})$. The reduction just multiplies each due date and each processing time by $2n\cdot\lceil\alpha\rceil$. This implies that the tardiness of each job is a multiple of $2n\cdot\lceil{\alpha}\rceil>\alpha\sum U_{j}$. Consequently, any schedule minimizing $T_{\max}+\alpha\sum U_{j}$ also minimizes $T_{\max}$. ∎

5 Lexicographically first minimizing $\sum U_{j}$ and then $T_{\max}$

We next prove the following theorem:

Theorem 4.

$1||Lex(\sum U_{j},T_{\max})$ is weakly NP-hard.

We will reduce from Partition:

Partition

Input:

A set of $n$ integers $a_{1},\ldots,a_{n}$.

Question:

Is there a subset $S\subset[n]$ such that $\sum_{i\in S}a_{i}=t$ where $t:=0.5\cdot\sum_{i=1}^{n}a_{i}$?

Somewhat similar to Section 3, there will be four jobs $J^{*}_{i}$, $\neg J^{*}_{i}$, $J_{i}$, and $\neg J_{i}$ for each $i\in[n]$. Again, scheduling $J^{*}_{i}$ and $J_{i}$ early will encode $a_{i}$ being part of the solution to Partition while $\neg J^{*}_{i}$ and $\neg J_{i}$ being early will encode $a_{i}$ not being part of the solution to Partition. However, while in Section 3 all these jobs had roughly the same processing time, this is not true anymore; in order to prevent that scheduling e.g. $J^{*}_{1}$ at the end allows to schedule one more early jobs, jobs $J_{i}$ and $\neg J_{i}$ will have much larger processing times than $J^{*}_{i}$ and $\neg J^{*}_{i}$, and the processing times will be increasing for increasing $i$. We add many small filler jobs $F_{i}$ to ensure that $J^{*}_{i}$ or $\neg J^{*}_{i}$ together with the filler jobs together has roughly the same size as $J_{i}$ and $\neg J_{i}$. Since all these filler jobs have the same characteristics, we will refer to each of these jobs as $F_{i}$; the set of all $2W/X$ filler jobs $F_{i}$ will be denoted as $\mathcal{F}_{i}$. The processing times of the filler jobs will be quite small, ensuring that any schedule with minimum number of tardy jobs schedules them early. We fix sufficiently large constants $W\gg X\gg Y\gg Z\gg t$ such that $W$ is a multiple of $X$ ($Z=(2t+1)$, $Y=(2t+1)\cdot Z$, $X=n\cdot 2^{n+2}\cdot Y$, and $W=2n^{2}\cdot X$ are possible choices).

For an overview of processing times and due dates, we refer to Table 2. We set the target maximum tardiness to $\ell:=n\cdot W+\sum_{i=1}^{n}i\cdot X+\sum_{i=1}^{n}2^{i}\cdot Y+t\cdot Z+t$.

Job	processing time	due date	multiplicity
$J^{*}_{i}$	$i\cdot X$	$i\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+t$	1
$\neg J^{*}_{i}$	$i\cdot X+a_{i}$	$(i-1)\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+t$	1
$F_{i}$	$X/2$	$i\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+t$	$2W/X$
$J_{i}$	$W+2^{i}\cdot Y+a_{i}\cdot Z$	$D_{1}^{*}+i\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+t\cdot Z+t$	1
$\neg J_{i}$	$W+2^{i}\cdot Y$	$D_{1}^{*}+i\cdot W+\sum_{i_{0}=1}^{i-1}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+t\cdot Z+t$	1

Table 2: Processing times and due dates where $D_{1}^{*}:=n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+t$.

We now prove correctness. We start with characterizing the set of early jobs of a schedule minimizing the number of tardy jobs. We start with the early jobs completed by time $D_{1}^{*}$:

Lemma 10.

Any set $\mathcal{J}^{*}$ of $n\cdot 2W/X+i$ early jobs completed by time $D_{1}^{*}$ has total processing time at least $n\cdot W+\sum_{{i_{0}}=1}^{i}{i_{0}}\cdot X$. If the total processing time of $\mathcal{J}^{*}$ is smaller than $n\cdot W+\sum_{{i_{0}}=1}^{i}{i_{0}}\cdot X+0.5\cdot X$, then $\mathcal{J}^{*}$ contains exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for each ${i_{0}}\in[i]$ as well as each filler job $F_{i_{0}}$.

Proof.

We first show that $\mathcal{J}^{*}$ does not contain $J_{i}$ and $\neg J_{i}$ for every $i\in[n]$. Each job has processing time at least $X/2$. Thus, if $\mathcal{J}^{*}$ would contain job $J_{i}$ or $\neg J_{i}$ which has processing time at least $W$, then

$$p(\mathcal{J}^{*})\geq(n\cdot 2W/X)\cdot X/2+W=(n+1)\cdot W>D_{1}^{*}\,,$$

a contradiction to all jobs form $\mathcal{J}^{*}$ finishing until $D_{1}^{*}$.

We prove the statement by induction on $i$. For $i=1$, note that each non-filler job has processing time at least $X$ and only $J^{*}_{1}$ and $\neg J^{*}_{1}$ have a processing time smaller than $2\cdot X$ while each filler job has processing time $X/2$. Thus, each set of $n\cdot 2W/X+1$ jobs has processing time at least $n\cdot W+X$. Further, each set of $n\cdot 2W/X+1$ early jobs containing at least two non-filler has processing time at least

$$(n\cdot 2W/X-1)\cdot(X/2)+2\cdot X=n\cdot W+1.5\cdot X,$$

and each set of $n\cdot 2W/X+1$ early jobs containing at least one non-filler job apart from $J_{1}^{*}$ and $\neg J_{1}^{*}$ has processing time at least

$$(n\cdot 2W/X)\cdot(X/2)+2\cdot X=n\cdot W+2\cdot X.$$

We continue with the induction step. Let $i>1$ and assume that the lemma holds for $i-1$. Let $J^{*}$ be the longest job from $\mathcal{J}^{*}$. As there are only $n\cdot 2W/X$ filler jobs and $\mathcal{J}^{*}$ contains no job from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[n]\}$ by our initial observation, $|\mathcal{J}^{*}|>n\cdot 2W/X$ implies that $J^{*}=J^{*}_{i_{1}}$ or $J^{*}=\neg J^{*}_{i_{1}}$ for some $i_{1}\in[n]$. We will now show that $i_{1}\geq i$, i.e., that the longest job from $\mathcal{J}^{*}$ has index at least $i$. So assume towards a contradiction that $i_{1}<i$, i.e., all jobs from $\mathcal{J}^{*}$ have index smaller than $i$. Since $|\mathcal{J}^{*}|=n\cdot 2W/X+i>n\cdot 2W/X+i_{1}$, this implies that there is some minimal $i_{2}$ such that $|\bigl{(}\{J^{*}_{i_{0}},\neg J^{*}_{i_{0}}:i_{0}\leq i_{2}\}\cup\mathcal{F}_{1}\cup\ldots\cup\mathcal{F}_{i_{2}}\bigr{)}\cap\mathcal{J}^{*}|>i_{2}\cdot(2W/X+1)$. By the minimality of $i_{2}$, we have $\{J^{*}_{i_{2}},\neg J^{*}_{i_{2}}\}\cup\mathcal{F}_{i_{2}}\subset\mathcal{J}^{*}$. Further, due to the minimality of $i_{2}$, we have that $p\Bigl{(}\bigl{(}\{J^{*}_{i_{0}},\neg J^{*}_{i_{0}}:i_{0}<i_{2}\}\cup\mathcal{F}_{1}\cup\ldots\cup\mathcal{F}_{i_{2}-1}\bigr{)}\cap\mathcal{J}^{*}\Bigr{)}\geq(i_{2}-1)\cdot W+\sum_{i_{0}=1}^{i_{2}-1}i_{0}\cdot X$. Consequently, we have

$$p\Bigl{(}\bigl{(}\{J^{*}_{i_{0}},\neg J^{*}_{i_{0}}:i_{0}\leq i_{2}\}\cup\mathcal{F}_{1}\cup\ldots\cup\mathcal{F}_{i_{2}}\bigr{)}\cap\mathcal{J}^{*}\Bigr{)} >(i_{2}-1)\cdot W+\sum_{i_{0}=1}^{i_{2}-1}i_{0}\cdot X+W+2\cdot i_{2}\cdot X$$

$$=i_{2}\cdot W+\sum_{i_{0}=1}^{i_{2}}i_{0}\cdot X+i_{2}\cdot X>d(F_{i_{2}})\,,$$

a contradiction to all jobs from $\mathcal{J}^{*}$ being early. Consequently, we have $p(J^{*})\geq i\cdot X$. By induction, $\mathcal{J}^{*}\setminus\{J^{*}\}$ has a total processing time of at least $n\cdot W+\sum_{{i_{0}}=1}^{i-1}{i_{0}}\cdot X$. Therefore, we have

$$p(\mathcal{J}^{*})=p(\mathcal{J}^{*}\setminus\{J^{*}\})+p(J^{*})\geq n\cdot W+\sum_{{i_{0}}=1}^{i-1}{i_{0}}\cdot X+i\cdot X=n\cdot W+\sum_{{i_{0}}=1}^{i}{i_{0}}\cdot X\,$$

and if $p(\mathcal{J}^{*})<n\cdot W+\sum_{{i_{0}}=1}^{i}{i_{0}}\cdot X+0.5\cdot X$, then $\mathcal{J}^{*}$ contains exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for each ${i_{0}}\in[i]$. ∎

Using Lemma 10, we now characterize the set of schedules with minimum number of early jobs (ignoring the secondary condition of minimizing the maximum tardiness for a moment). In order to do so, we introduce EDD-schedules, a class of schedules used to solve $1||\sum w_{j}U_{j}$. An EDD-schedule for a set $\widetilde{\mathcal{J}}$ is a schedule which schedules all jobs from $\widetilde{\mathcal{J}}$ by non-decreasing order of their due dates, and then all other jobs in arbitrary order afterwards. It is a well-known fact that if there is a schedule where a set $\widetilde{\mathcal{J}}$ of jobs is early, then the EDD-schedule for $\widetilde{\mathcal{J}}$ is such a schedule (see e.g. [1]). Note that if $\ell>P$, then the canonical schedule for $\widetilde{\mathcal{J}}$ (as defined in Section 2) is an EDD-schedule for $\widetilde{\mathcal{J}}$.

Lemma 11.

A schedule minimizes the number of tardy jobs if and only if the set $\widetilde{\mathcal{J}}$ of early jobs consists of

• all filler jobs,

• for every $i\in[n]$, exactly one of $J^{*}_{i}$ and $\neg J^{*}_{i}$,

• $\sum_{i\in[n]:\neg J^{*}_{i}\in\widetilde{\mathcal{J}}}a_{i}\leq t$, and

• exactly $n$ jobs from $\{J_{i},\neg J_{i}:i\in[n]\}$ such that for each $i\in[n]$, at most $i$ jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[i]\}$ are early.

Proof.

We first show that for set $\mathcal{\widetilde{J}}$ fulfilling the four bullet points, the EDD-schedule $\sigma$ for $\widetilde{\mathcal{J}}$ schedules every job from $\mathcal{\widetilde{J}}$ early. We start with the filler jobs and jobs $J_{i}^{*}$. Until $F_{i}$ or $J_{i}^{*}$, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\leq i$ and $\mathcal{F}_{i_{0}}$ for $i_{0}\leq i$ can be scheduled. Thus, $F_{i}$ or $J_{i}^{*}\in\mathcal{\widetilde{J}}$ is completed by time

$$\sum_{i_{0}\in[i]:J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(J_{i_{0}}^{*})+\sum_{i_{0}\in[i]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(\neg J_{i_{0}}^{*})+\sum_{i_{0}=1}^{i}p(\mathcal{F}_{i_{0}}) =\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}\in[i]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}a_{i_{0}}+i\cdot W$$

$$\leq i\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+t=d(F_{i})=d(J_{i}^{*})$$

using the third bullet point for the inequality.

We continue with job $\neg J_{i}^{*}\in\mathcal{\widetilde{J}}$. Until $\neg J_{i}^{*}$, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\leq i$ and $\mathcal{F}_{i_{0}}$ for $i_{0}<i$ is scheduled. Thus, $\neg J_{i}^{*}\in\mathcal{\widetilde{J}}$ is completed by time

$$\sum_{i_{0}\in[i]:J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(J_{i_{0}}^{*})+\sum_{i_{0}\in[i]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(\neg J_{i_{0}}^{*})+\sum_{i_{0}=1}^{i-1}p(\mathcal{F}_{i_{0}}) =\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}\in[i]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}a_{i_{0}}+(i-1)\cdot W$$

$$\leq(i-1)\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+t=d(\neg J_{i}^{*})$$

using again the third bullet point for the inequality.

Finally, we consider job $J_{i}\in\mathcal{\widetilde{J}}$ or $\neg J_{i}\in\mathcal{\widetilde{J}}$. Until $J_{i}$ or $\neg J_{i}$, all filler jobs, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$, and up to $i$ jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[i]\}$ (by the fourth bullet point) are scheduled. Thus, the job is completed by time

$$\sum_{i_{0}=1}^{n}\bigl{(}p(\mathcal{F}_{i_{0}})+p(\neg J_{i}^{*})\bigr{)}+\sum_{J\in\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[i]\}\cap\widetilde{\mathcal{J}}}p(J) =n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+2t+\sum_{J\in\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[i]\}\cap\widetilde{\mathcal{J}}}p(J)$$

$$\leq D_{1}^{*}+t+i\cdot W+i\cdot 2^{i_{0}}\cdot Y+\sum_{i_{0}:J_{i_{0}}\in\widetilde{\mathcal{J}}}a_{i_{0}}\cdot Z$$

$$<D_{1}^{*}+i\cdot W+X\,.$$

Note that this last term is smaller than $d(J_{i})$ and $d(\neg J_{i})$ except for $\neg J_{1}$. For $\neg J_{1}$, note that $\{J_{i_{0}}:i_{0}\in[i]\}\cap\widetilde{\mathcal{J}}=\emptyset$ and thus, $\neg J_{1}$ is completed by time $D_{1}^{*}+t+W+2\cdot Y<d(\neg J_{1})$.

We continue with the backwards direction. We show that any schedule has at most $n\cdot(2W/X)+2n$ early jobs, and equality only holds if the four bullet points hold. By Lemma 10, there are at most $2n\cdot W/X+n$ early jobs completed by time $D_{1}^{*}$. We continue by analyzing the early jobs completed after $D_{1}^{*}$. The only jobs with due date larger than $D_{1}^{*}$ are $\{J_{i},\neg J_{i}:i\in[n]\}$, all of which have processing time at least $W$. Since the largest due date is $d(J_{n})<D_{1}^{*}+(n+1)\cdot W$, it follows that at most $n$ jobs completed after $D_{1}^{*}$ are early. As there is a schedule with $2n\cdot W/X+2n$ early jobs (as shown in the forward direction), this implies that the fourth bullet point holds.

We continue by showing the first two bullet points. Note that $D_{1}^{*}<n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+0.5\cdot X$. Lemma 10 then implies that all filler jobs and exactly one of $J^{*}_{i}$ and $\neg J^{*}_{i}$ for $i\in[n]$ is early.

It remains to show the third bullet point. We may assume, without loss of generality and due to the EDD order, that the last early job from $\{J_{i}^{*},\neg J_{i}^{*}:i\in[n]\}\cup\mathcal{F}_{1}\cup\ldots\cup\mathcal{F}_{n}$ is $F_{n}$. Then $F_{n}$ is completed by time

$$\sum_{i_{0}\in[n]:J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(J_{i_{0}}^{*})+\sum_{i_{0}\in[n]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}p(\neg J_{i_{0}}^{*})+\sum_{i_{0}=1}^{n}p(\mathcal{F}_{i_{0}}) =\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}\in[n]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}a_{i_{0}}+n\cdot W$$

$$=d(F_{n})+\sum_{i_{0}\in[n]:\neg J^{*}_{i_{0}}\in\mathcal{\widetilde{J}}}a_{i_{0}}-t\,.$$

Because $F_{n}$ is early, this implies the third bullet point. ∎

Since Lemma 11 implies that the minimum number of tardy jobs is $2n$ in any schedule, we will use $k:=2n$ in the following. If we require that the maximum tardiness is at most $\ell$ in addition to the schedule having only $k$ tardy jobs, then we get also structural restrictions on which jobs from $\{J_{i},\neg J_{i}:i\in[n]\}$ are early:

Lemma 12.

Any canonical schedule with minimum number of tardy jobs and tardiness at most $\ell$ schedules exactly one of $J_{i}$ and $\neg J_{i}$ for each $i\in[n]$.

Proof.

By Lemma 11, there are $n$ early jobs from $\{J_{i},\neg J_{i}:i\in[n]\}$. Assume towards a contradiction that the statement is not true, i.e., there exists a canonical schedule $\sigma$ and some $i\in[n]$ such that neither $J_{i}$ nor $\neg J_{i}$ or both $J_{i}$ and $\neg J_{i}$ are early. First consider the case that there exist some $i$ so that both $J_{i}$ and $\neg J_{i}$ are early and there are at least $i-1$ early jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}<i\}$. By Lemma 11, all filler jobs are early. Since $d(F_{i_{0}})<d(\neg J_{i})<d(J_{i})$ for every $i_{0}\in[n]$, every filler job $F_{i_{0}}$ is completed before $\neg J_{i}$ and $J_{i}$ since $\sigma$ is canonical. Consequently, $J_{i}$ is completed not before time

$$n\cdot W+(i-1)\cdot W+W+W=(n+i+1)\cdot W>d(J_{i})\,,$$

a contradiction to $J_{i}$ being early.

We now consider the other case, for any $i\in[n]$ such that $J_{i}$ and $\neg J_{i}$ are early, there are at most $i-2$ early jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}<i\}$. Consequently, there exist some $i$ such that both $J_{i}$ and $\neg J_{i}$ are early and for all $i_{1}>i$, exactly one of $J_{i_{1}}$ and $\neg J_{i_{1}}$ is early. In particular, at least one of $J_{n}$ and $\neg J_{n}$ is early (both are early if $i=n$). If $J_{n}$ is early, then when $J_{n}$ is completed, the following jobs are completed as well: all filler jobs, jobs $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\in[n]$ (because they have tardiness at most $\ell$ by assumption and $d(\neg J^{*}_{i_{0}})+\ell<d(J^{*}_{i_{0}})+\ell<d(J_{i})$), $i-2$ jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}<i\}$, $J_{i}$ and $\neg J_{i}$, and exactly one of $J_{i_{0}}$ and $\neg J_{i_{0}}$ for $i_{0}>i$. Therefore, $J_{n}$ is completed not before time

$$n\cdot W+2\cdot\sum_{i_{0}=1}^{n}i_{0}\cdot X+t+n\cdot W+2\cdot 2^{i}\cdot Y+\sum_{i_{0}=i+1}^{n}2^{i_{0}}\cdot Y>d(J_{n})\,,$$

a contradiction to $J_{n}$ being early. If $\neg J_{n}$ is early, then an analogous calculation leads to a contradiction. ∎

We call a set $\widetilde{\mathcal{J}}$ fulfilling the conditions of Lemmas 11 and 12 (that is, containing every filler job and for each $i\in[n]$, exactly one of $J^{*}_{i}$ and $\neg J^{*}_{i}$ as well as exactly one of $J_{i}$ and $\neg J_{i}$, and fulfilling $\sum_{i:\neg J_{i}^{*}\in\mathcal{\widetilde{J}}}a_{i}\leq t$) a candidate set. By Lemma 11, we may restrict ourselves to schedules $\sigma$ constructed by applying Observation 1 to a candidate set. We will further assume that ties between $J^{*}_{i}$ and $F_{i}$ will be broken in favor of $J^{*}_{i}$ (i.e., if $\widetilde{d}_{\sigma}(J^{*}_{i})=\widetilde{d}_{\sigma}(F_{i})$, then $J^{*}_{i}$ will be scheduled before $F_{i}$), and a possible tie between $J^{*}_{n}$ and $J_{n}$ will be broken in favor of $J_{n}$. Thus, given a candidate set $\widetilde{\mathcal{J}}$, the corresponding schedule looks as follows:

1. First, $i=1$ to $n$, job $J^{*}_{i}$ if $J^{*}_{i}\in\widetilde{\mathcal{J}}$ or $\neg J^{*}_{i}$ otherwise (i.e., $\neg J^{*}_{i}\in\widetilde{\mathcal{J}}$). In both cases, $\mathcal{F}_{i}$ follows.

2. Second, $i=1$ to $n$, the following jobs are scheduled: If $J^{*}_{i}\in\widetilde{\mathcal{J}}$, then $\neg J^{*}_{i}$ is scheduled, followed by $J_{i}$ if $J_{i}\in\widetilde{\mathcal{J}}$ or $\neg J_{i}$ if $\neg J_{i}\in\widetilde{\mathcal{J}}$. Otherwise (i.e., if $J^{*}_{i}\notin\widetilde{\mathcal{J}}$), we first schedule $J_{i}$ if $J_{i}\in\widetilde{\mathcal{J}}$ or $\neg J_{i}$ if $\neg J_{i}\in\widetilde{\mathcal{J}}$ and afterwards schedule $J^{*}_{i}$.

3. Third, for $i=1$ to $n$, we schedule $J_{i}$ if $\neg J_{i}\in\widetilde{\mathcal{J}}$ or $\neg J_{i}$ if $J_{i}\in\widetilde{\mathcal{J}}$.

An example of the schedule for a candidate set can be found in Example 2.

Example 2.

Assume that $n=4$ and $\widetilde{\mathcal{J}}=\{J^{*}_{1},\neg J^{*}_{2},J^{*}_{3},\neg J^{*}_{4}\}\cup\{J_{1},\neg J_{2},J_{3},\neg J_{4}\}\cup\{\mathcal{F}_{1},\mathcal{F}_{2},\mathcal{F}_{3},\mathcal{F}_{4}\}$. Then the schedule for $\widetilde{\mathcal{J}}$ is

$$\begin{array}[]{rrrrrrrr}J^{*}_{1},&\mathcal{F}_{1},&\neg J^{*}_{2},&\mathcal{F}_{2},&J^{*}_{3},&\mathcal{F}_{3},&\neg J^{*}_{4},&\mathcal{F}_{4},\\ \neg J^{*}_{1},&J_{1},&\neg J_{2},&J^{*}_{2},&\neg J^{*}_{3},&J_{3},&\neg J_{4},&J^{*}_{4},\\ \neg J_{1},&J_{2},&\neg J_{3},&J_{4}.\end{array}$$

For a candidate set $\widetilde{\mathcal{J}}$, we define $\widetilde{S}^{*}:=\{i\in[n]:J^{*}_{i}\in\widetilde{\mathcal{J}}\}$ and $\widetilde{S}:=\{i\in[n]:J_{i}\in\widetilde{\mathcal{J}}\}$.

We now want to characterize when a candidate set $\mathcal{\widetilde{J}}$ corresponds to a solution to the constructed instance of $1||Lex(\sum U_{j},T_{\max})$. We first show that all filler jobs as well as jobs $J_{i}^{*}$ or $\neg J_{i}^{*}$ from $\mathcal{\widetilde{J}}$ are early:

Lemma 13.

Let $\widetilde{\mathcal{J}}$ be a candidate set. Then all jobs from $J\in\mathcal{J}^{*}:=\widetilde{\mathcal{J}}\cap\bigl{(}\{J^{*}_{i},\neg J^{*}_{i}\}\cup\mathcal{F}_{i}\bigr{)}$ are early.

Proof.

Note that $\ell>D_{1}^{*}\geq d(J)$ for every $J\in\mathcal{J}^{*}:=\widetilde{\mathcal{J}}\cap\bigl{(}\{J^{*}_{i},\neg J^{*}_{i}\}\cup\mathcal{F}_{i}\bigr{)}$. Therefore, these jobs are the first jobs in the candidate schedule $\sigma$ for $\mathcal{\widetilde{J}}$, implying that they are early in $\sigma$ if and only if they are early in the EDD-schedule for $\mathcal{\widetilde{J}}$ which holds by Lemma 11. ∎

Next, we analyze when the remaining jobs are early, i.e., the jobs from $\widetilde{\mathcal{J}}\cap\{J_{i},\neg J_{i}:i\in[n]\}$.

Lemma 14.

Let $\widetilde{\mathcal{J}}$ be a candidate set. Then all $n$ jobs from $\widetilde{\mathcal{J}}\cap\{J_{i},\neg J_{i}:i\in[n]\}$ are early if and only if

• for each $i\in[n]$ with $J^{*}_{i}\in\widetilde{\mathcal{J}}$, we have $J_{i}\in\widetilde{\mathcal{J}}$ and $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\leq t$, and

• for each $i\in[n]$ such that $\neg J_{i},\neg J_{i}^{*}\in\mathcal{\widetilde{J}}$, we have $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\leq t$.

Proof.

First consider the case that $J^{*}_{i}\in\widetilde{\mathcal{J}}$. By the definition of a candidate set, exactly one of $J_{i}$ and $\neg J_{i}$ is contained in $\mathcal{\widetilde{J}}$. We will now show that job $\neg J_{i}$ will be tardy even if $\neg J_{i}\in\mathcal{\widetilde{J}}$ while $J_{i}$ is tardy if and only if $J_{i}\notin\mathcal{\widetilde{J}}$ and $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}>t$. Job $J_{i}$ if $J_{i}\in\widetilde{\mathcal{J}}$ or $\neg J_{i}$ if $\neg J_{i}\in\widetilde{\mathcal{J}}$ is completed after all filler jobs, jobs $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\in[i]$, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}>i$, $J_{i_{0}}$ for $i_{0}\in[i]\cap\widetilde{S}$, and $\neg J_{i_{0}}$ for $i_{0}\in[i]\setminus\widetilde{S}$. To simplify the following equations, we assume that for each $i_{0}>i$, job $\neg J^{*}_{i_{0}}$ and not $J^{*}_{i_{0}}$ is scheduled before $J_{i}$ (this increases the total completion time by at most $2t<Z$). Thus, $J_{i}$ or $\neg J_{i}$ is completed by time

$$\sum_{i_{0}=1}^{n}p(\mathcal{F}_{i_{0}})+ \sum_{i_{0}=1}^{i}\Bigl{(}p(J^{*}_{i_{0}})+p(\neg J^{*}_{i_{0}})\Bigr{)}+\sum_{i_{0}=i+1}^{n}p(\neg J^{*}_{i_{0}})+\sum_{i_{0}\in[i]\cap\widetilde{S}}p(J_{i_{0}})+\sum_{i_{0}\in[i]\setminus\widetilde{S}}p(\neg J_{i_{0}})$$

$$=n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i}i_{0}\cdot X+2t+i\cdot W+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z$$

$$=(n+i)\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z+2t$$

$$=D_{1}^{*}+i\cdot W+\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z+t$$

$$=d(J_{i})+\bigl{(}\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z$$

$$=d(\neg J_{i})+i\cdot X+\bigl{(}\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z\,.$$

Thus, if $\neg J_{i}\in\widetilde{\mathcal{J}}$, then $\neg J_{i}$ is tardy. If $J_{i}\in\widetilde{\mathcal{J}}$, then $J_{i}$ is early if and only if $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\leq t$.

We continue with the case $\neg J^{*}_{i}\in\widetilde{\mathcal{J}}$. We will show that $J_{i}$ is early if and only if $J_{i}\in\mathcal{\widetilde{J}}$ while $\neg J_{i}$ is early if and only if $\neg J_{i}\in\mathcal{\widetilde{J}}$ and $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\leq t$. Job $J_{i}$ if $J_{i}\in\widetilde{\mathcal{J}}$ or $\neg J_{i}$ if $\neg J_{i}\in\widetilde{\mathcal{J}}$ is completed after all filler jobs, jobs $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\in[i-1]$, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\geq i$, $J_{i_{0}}$ for $i\in[i]\cap\widetilde{S}$, and $\neg J_{i_{0}}$ for $i\in[i]\setminus\widetilde{S}$. As in the case $J^{*}_{i}\in\widetilde{\mathcal{J}}$, we assume that for each $i_{0}\geq i$, job $\neg J^{*}_{i_{0}}$ and not $J^{*}_{i_{0}}$ is scheduled before $J_{i}$ (this increases the total completion time by at most $2t<Z$). Thus, $J_{i}$ or $\neg J_{i}$ is completed by time

$$\sum_{i_{0}=1}^{n}p(\mathcal{F}_{i_{0}})+ \sum_{i_{0}=1}^{i-1}\Bigl{(}p(J^{*}_{i_{0}})+p(\neg J^{*}_{i_{0}})\Bigr{)}+\sum_{i_{0}=i}^{n}p(\neg J^{*}_{i_{0}})+\sum_{i_{0}\in[i]\cap\widetilde{S}}p(J_{i_{0}})+\sum_{i_{0}\in[i]\setminus\widetilde{S}}p(\neg J_{i_{0}})$$

$$=(n+i)\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i-1}i_{0}\cdot X+2t+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z$$

$$=D_{1}^{*}+i\cdot W+\sum_{i_{0}=1}^{i-1}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z+t$$

$$=d(J_{i})-i\cdot X+\bigl{(}\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z$$

$$=d(\neg J_{i})+\bigl{(}\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z$$

Therefore, $J_{i}$ is always early while $\neg J_{i}$ is early if and only if $\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\leq t$.

Since a candidate job contains exactly one of $J_{i}$ and $\neg J_{i}$ for every $i\in[n]$, $n$ jobs from $\mathcal{\widetilde{J}}$ are early if and only if exactly one of $J_{i}$ and $\neg J_{i}$ is early for every $i\in[n]$. As shown above, this is in turn equivalent to the two bullet points. ∎

Finally, we analyze when the tardiness is at most $\ell$.

Lemma 15.

Let $\widetilde{\mathcal{J}}$ be a candidate set. Then the maximum tardiness is at most $\ell$ if and only if $J^{*}_{n}\in\widetilde{\mathcal{J}}$ or $\sum_{i\in\widetilde{S}}a_{i}\leq t$.

Proof.

We make a distinction on the different jobs. First consider a job $J^{*}_{i}$, and assume that $J^{*}_{i}\notin\widetilde{\mathcal{J}}$ (otherwise $J^{*}_{i}$ will be completed earlier than the time we compute below). Until $J^{*}_{i}$, the following jobs are scheduled: all filler jobs, jobs $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\leq i$, exactly one of $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}>i$, and $J_{i_{0}}$ for $i_{0}\in[i]\cap\widetilde{S}$, and $\neg J_{i_{0}}$ for $i_{0}\in[i]\setminus\widetilde{S}$. Consequently, $J^{*}_{i}$ is completed by time

$$\sum_{i_{0}=1}^{n}p(\mathcal{F}_{i_{0}})+ \sum_{i_{0}=1}^{i}\Bigl{(}p(J^{*}_{i_{0}})+p(\neg J^{*}_{i_{0}})\Bigr{)}+\sum_{i_{0}=i+1}^{n}p(\neg J^{*}_{i_{0}})+\sum_{i_{0}\in[i]\cap\widetilde{S}}p(J_{i_{0}})+\sum_{i_{0}\in[i]\setminus\widetilde{S}}p(\neg J_{i_{0}})$$

$$=n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}=1}^{n}a_{i_{0}}+i\cdot W+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z$$

$$\leq(n+i)\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z+2t$$

$$=d(J^{*}_{i})+n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}\cdot Z+t$$

$$=d(J^{*}_{i})+\ell-\sum_{i_{0}=i+1}^{n}2^{i_{0}}\cdot Y+\bigl{(}\sum_{i_{0}\in[i]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z-t$$

which is smaller than $d(J^{*}_{i})+\ell$ for $i<n$ or $\sum_{i\in\widetilde{S}}a_{i_{0}}\leq t$.

For $\neg J^{*}_{i}$, the calculations are identical to the one for $J^{*}_{i}$, except that neither $J_{i}$ nor $\neg J_{i}$ is scheduled before $\neg J^{*}_{i}$ and the due date of $J^{*}_{i}$ is by $W$ smaller. Consequently, $\neg J^{*}_{i}$ is completed by time

$$\sum_{i_{0}=1}^{n}p(\mathcal{F}_{i_{0}})+ \sum_{i_{0}=1}^{i}\Bigl{(}p(J^{*}_{i_{0}})+p(\neg J^{*}_{i_{0}})\Bigr{)}+\sum_{i_{0}=i+1}^{n}p(\neg J^{*}_{i_{0}})+\sum_{i_{0}\in[i-1]\cap\widetilde{S}}p(J_{i_{0}})+\sum_{i_{0}\in[i-1]\setminus\widetilde{S}}p(\neg J_{i_{0}})$$

$$=d(\neg J^{*}_{i})+n\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{i-1}2^{i_{0}}\cdot Y+\sum_{i_{0}\in[i-1]\cap\widetilde{S}}a_{i_{0}}\cdot Z+t$$

$$=d(\neg J^{*}_{i})+\ell-\sum_{i_{0}=i}^{n}a_{i_{0}}\cdot Y+\bigl{(}\sum_{i_{0}\in[i-1]\cap\widetilde{S}}a_{i_{0}}-t\bigr{)}\cdot Z-t\,.$$

Thus, $\neg J^{*}_{i}$ always has tardiness smaller than $\ell$.

For the filler job $F_{i}$, note that $F_{i}$ is completed before $J^{*}_{i}$ for $J^{*}_{i}\notin\widetilde{\mathcal{J}}$ but has the same due date. Thus, $F_{i}$ has tardiness smaller than $\ell$.

We continue with jobs $J_{i}\notin\widetilde{\mathcal{J}}$ or $\neg J_{i}\notin\widetilde{\mathcal{J}}$. This job is completed after all filler jobs, jobs $J^{*}_{i_{0}}$ and $\neg J^{*}_{i_{0}}$ for $i_{0}\in[n]$, and $(n+i)$ jobs from $\{J_{i_{0}},\neg J_{i_{0}}:i_{0}\in[n]\}$. Consequently, the job is completed by time

$$n\cdot W+ 2\cdot\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{n}a_{i_{0}}+(n+i)\cdot W+\sum_{i_{0}=1}^{n}2^{i_{0}}\cdot Y+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+\sum_{i_{0}=1}^{n}a_{i_{0}}\cdot Z$$

$$=D_{1}^{*}+(n+i)\cdot W+\sum_{i_{0}=1}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{n}2^{i_{0}}\cdot Y+\sum_{i_{0}=1}^{i}2^{i_{0}}\cdot Y+2t\cdot Z+t$$

$$=d(\neg J_{i})+n\cdot W+\sum_{i_{0}=i}^{n}i_{0}\cdot X+\sum_{i_{0}=1}^{n}2^{i_{0}}\cdot Y+t\cdot Z$$

$$<d(\neg J_{i})+\ell<d(J_{i})+\ell\,.$$

Thus, jobs $J_{i}$ and $\neg J_{i}$ have tardiness smaller than $\ell$. ∎

We can now show the correctness of the reduction:

Proof of Theorem 4.

We start with the forward direction, so assume that there is a solution $S$ to the Partition instance. Let $\widetilde{\mathcal{J}}:=\{J^{*}_{i},J_{i}:i\in S\}\cup\{\neg J^{*}_{i},\neg J_{i}:i\notin S\}\cup\mathcal{F}_{1}\cup\ldots\cup\mathcal{F}_{n}$. By Lemmas 14 and 13, the schedule for $\widetilde{\mathcal{J}}$ has $k$ tardy jobs. By Lemma 15, the schedule has tardiness smaller than $\ell$.

We continue with the reverse direction. So assume that there is a schedule $\sigma$ with $k$ tardy jobs and tardiness at most $\ell$. By Lemmas 11 and 12, we may assume that $\sigma$ is the candidate schedule for some candidate set $\widetilde{\mathcal{J}}$. By Lemma 13, we have $\sum_{i\in\widetilde{S}^{*}}a_{i}\geq t$. Further, we have $\sum_{i\in\widetilde{S}}a_{i}\leq t$: If $J^{*}_{n}\in\widetilde{\mathcal{J}}$, then this follows from Lemma 14. Otherwise we have $J^{*}_{n}\notin\widetilde{\mathcal{J}}$ and the inequality follows from Lemma 15. By Lemma 15, we have $\widetilde{S}^{*}\subseteq\widetilde{S}$. Combining these statements, we get

$$t\leq\sum_{i\in\widetilde{S}^{*}}a_{i}\leq\sum_{i\in\widetilde{S}}a_{i}\leq t\,.$$

Consequently, we have $\widetilde{S}^{*}=\widetilde{S}$, and the inequalities hold with equality. Therefore, $\widetilde{S}^{*}$ is a solution to the Partition instance. ∎

6 Conclusions

In this paper we resolved one of the most fundamental problems in bicriteria scheduling which involves combining the maximal tardiness objective with the total number of tardy jobs objective. We proved that the lexicographic version of this problem (i.e. the $1||Lex(T_{\max},\sum U_{j})$ problem) is strongly NP-hard when the maximal tardiness is used as the primary criterion, while it is at least weakly NP-hard when the number of tardy jobs is used as the primary criterion (i.e. the $1||Lex(\sum U_{j},T_{\max})$ problem). We also classified the two other variants of the problem, $1|T_{\max}\leq\ell,\sum U_{j}\leq k|$ and $1||\alpha T_{\max}+\sum U_{j}$, as strongly NP-hard.

The first obvious question raised by our work is whether $1||Lex(\sum U_{j},T_{\max})$ is strongly NP-hard or not. However, this is only one out of several other objectives that may be considered in future work. For example, there is lack of practical (polynomial time) approximation algorithms for solving bicriteria scheduling problems in general. It is interesting to see if one can design a polynomial time approximation scheme (PTAS) to solve our specific bicriteria problem, and if not maybe one can rule out the existence of such an algorithm. Another very interesting research direction which is still unexplored is whether we can provide FPT algorithms to hard multicriteria scheduling problems with respect to some of the more natural parameters, e.g., the number of different due dates or the number of different processing times in the instance.