Combinatorial n-fold Integer Programming and Applications

TL;DR

This paper introduces a single-exponential algorithm for combinatorial n-fold integer programs, enabling faster solutions for various NP-hard problems by allowing variable dimensions and improving over traditional ILP methods.

Contribution

The authors develop a novel single-exponential algorithm for combinatorial n-fold integer programs that handles variable dimensions and applies augmentation techniques, improving efficiency over Lenstra's algorithm.

Findings

Achieved exponential speedups for problems like Closest String and Swap Bribery.

Demonstrated the effectiveness of augmentation techniques in combinatorial n-fold IPs.

Provided new insights into problem structures through the existence of local augmenting steps.

Abstract

Many fundamental NP-hard problems can be formulated as integer linear programs (ILPs). A famous algorithm by Lenstra solves ILPs in time that is exponential only in the dimension of the program, and polynomial in the size of the ILP. That algorithm became a ubiquitous tool in the design of fixed-parameter algorithms for NP-hard problems, where one wishes to isolate the hardness of a problemby some parameter. However, in many cases using Lenstra's algorithm has two drawbacks: First, the run time of the resulting algorithms is often doubly-exponential in the parameter, and second, an ILP formulation in small dimension cannot easily express problems involving many different costs. Inspired by the work of Hemmecke, Onn and Romanchuk [Math. Prog. 2013], we develop a single-exponential algorithm for so-called combinatorial n-fold integer programs, which are remarkably similar to prior ILP formulations for various problems, but unlike them, also allow variable dimension. We then apply our algorithm to a few representative problems like Closest String, Swap Bribery, Weighted Set Multicover, and obtain exponential speedups in the dependence on the respective parameters, the input size, or both. Unlike Lenstra's algorithm, which is essentially a bounded search tree algorithm, our result uses the technique of augmenting steps. At its heart is a deep result stating that in combinatorial n-fold IPs, existence of an augmenting step implies existence of a \local" augmenting step, which can be found using dynamic programming. Our results provide an important insight into many problems by showing that they exhibit this phenomenon, and highlights the importance of augmentation techniques.

Tables

Table 1. Table 1: Run time improvements for a few representative problems resulting from this work.

Problem	Previous best run time	Our result
Closest String	$2^{2^{O\left(k\log k\right)}}O\left(\log L\right)$ [GNR03]	$k^{O\left(k^2\right)}O\left(\log L\right)$
Optimal Consensus	$\mathrm{𝖥𝖯𝖳}$ for $k\le 3$, open for $k\ge 4$ [ALN+11]	$k^{O\left(k^2\right)}O\left(\log L\right)$
Score-Swap Bribery	$2^{2^{O\left(\left|C\right|\log \left|C\right|\right)}}O\left(\log \left|V\right|\right)$ [DS12]	${\left|C\right|}^{O\left({\left|C\right|}^2\right)}O\left(T^3\log \left|V\right|\right)$,
	${\left|C\right|}^{O\left({\left|C\right|}^6\right)}O\left({\left|V\right|}^3\right)$ [KKM17b]	with $T\le \left|V\right|$
C1-Swap Bribery	$2^{2^{O\left(\left|C\right|\log \left|C\right|\right)}}O\left(\log \left|V\right|\right)$ [DS12]	${\left|C\right|}^{O\left({\left|C\right|}^4\right)}O\left(T^3\log \left|V\right|\right)$,
	${\left|C\right|}^{O\left({\left|C\right|}^6\right)}O\left({\left|V\right|}^3\right)$ [KKM17b]	with $T\le \left|V\right|$
Weighted Set Multicover	$2^{2^{O\left(k\log k\right)}}O\left(n\right)$ [BFN+15]	$k^{O\left(k^2\right)}O\left(\log n\right)$
Huge $n$-fold IP	$\mathrm{𝖥𝖯𝖳}$ with $D=I$ and	$\mathrm{𝖥𝖯𝖳}$ with parameter-
	$A$ totally unimodular	sized domains

Table 2. Table 2: If the “specialization” row does not contain a value, it means its “default” value is assumed. The default values are δ = d H 𝛿 subscript 𝑑 𝐻 \delta=d_{H} , b = 0 𝑏 0 b=0 , and for all i ∈ [ k ] 𝑖 delimited-[] 𝑘 i\in[k] , d i = 0 subscript 𝑑 𝑖 0 d_{i}=0 , D i = L subscript 𝐷 𝑖 𝐿 D_{i}=L . In each row corresponding to a problem, the last column gives the run time of the algorithm with the slowest-growing dependency on k 𝑘 k . Most problems either reduce to Closest String and thus derive their time complexity from the result of Gramm et al. even though the original paper does not mention these problems; in that case, we write  [ GNR03 , implicit] . For some problems, no fixed-parameter algorithm was known before, but it is not difficult to see that the ILP formulation of Gramm et al. could be modified to model these problems as well; in that case, we write  [ GNR03 , “implicit”] .

Problem	Specialization of $\delta$-Multi Strings	Previous best run time / hardness
Closest String	$D_i=d$ for all $i\in \left[k\right]$	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03]
Farthest String	$d_i=d$ for all $i\in \left[k\right]$	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03, “implicit”]
$d$-Mismatch	poly-reduces to Closest String [GNR03]	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03]
Distinguishing String Selection	special case of Neighbor String	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03, “implicit”]
Neighbor String	$D_i=d_i$ for all $i\in \left[k\right]$	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03, “implicit”]
Closest String with Wildcards	$D_i=d$ for all $i\in \left[k\right]$	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [HR15]
Closest to Most Strings	$\mathrm{𝖥𝖯𝖳}$-reduces to Closest String [BM11]	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03, implicit]
$c$-HRC	$\mathrm{𝖥𝖯𝖳}$-reduces to Closest String [AFRS14]	$2^{2^{O\left(k\log k\right)}}{\log }^{O\left(1\right)}n$ [GNR03, implicit]
Optimal Consensus	$D_i=d$ for all $i\in \left[k\right]$, $b=1$	$\mathrm{𝖥𝖯𝖳}$ for $k=3$, open for $k>3$ [ALN+11]

Equations

$$\begin{array}[]{rcl|rclr}\displaystyle\sum_{e\in[k]}\sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}\in[k]^{k}}d_{H}(e,f_{j})x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e}&\leq&d&\displaystyle\sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}\in[k]^{k}}\sum_{({\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime},e)\in[k]^{k+1}}d_{H}(e,f_{j})x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime},e}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&\leq&d&\forall j\in[k]\\ \displaystyle\sum_{e\in[k]}x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e}&=&b^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&\displaystyle\sum_{({\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime},e)\in[k]^{k+1}}x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime},e}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&=&b^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&\forall{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}\in[k]^{k}\\ x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e}&\geq&0&&&&\forall({\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e)\in[k]^{k+1}\\ &&&x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime}}&=&0&\forall{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}^{\prime}\neq{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},\forall e\in[k]\\ &&&0\leq x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}},e}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&\leq&b^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}&\forall{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}\in[k]^{k}\end{array}$$

$$\begin{array}[]{ccccl|cccccl}D_{1}&D_{2}&\cdots&D_{k^{k}}&\leq d&\nobreak\leavevmode\hfil&D&D&\cdots&D&\leq d\\ \mathbf{1}^{\intercal}&0&\cdots&0&=b^{1}&&\mathbf{1}^{\intercal}&0&\cdots&0&=b^{1}\\ 0&\mathbf{1}^{\intercal}&\cdots&0&=b^{2}&&0&\mathbf{1}^{\intercal}&\cdots&0&=b^{2}\\ \vdots&\vdots&\ddots&\vdots&=\vdots&&\vdots&\vdots&\ddots&\vdots&=\vdots\\ 0&0&\cdots&\mathbf{1}^{\intercal}&=b^{k^{k}}&&0&0&\cdots&\mathbf{1}^{\intercal}&=b^{k^{k}},\\ \end{array}$$

$$\displaystyle\min\left\{f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\,\mid E^{(n)}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}={\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}}\,,\ {\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}\,,\ {\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\in\mathbb{Z}^{nt}\right\},$$

$$\displaystyle\leavevmode\nobreak\ \leavevmode\nobreak\ \mbox{where }\leavevmode\nobreak\ E^{(n)}:=\left(\begin{array}[]{cccc}D&D&\cdots&D\\ A&0&\cdots&0\\ 0&A&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&A\\ \end{array}\right).$$

$$A=(1\text{leavevmode}\cdots \text{leavevmode}1)$$

$$\min\left\{f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\mid E^{(n)}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}={\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}},{\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}},{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\in\mathbb{Z}^{nt}\right\}\enspace.$$

$$\min\left\{f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\,\mid A{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}={\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}}\,,\ {\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}\,,\ {\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\in\mathbb{Z}^{n}\right\}.$$

$$g(E)\leq\max_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf v$}}{\mbox{\boldmath$\textstyle\bf v$}}{\mbox{\boldmath$\scriptstyle\bf v$}}{\mbox{\boldmath$\scriptscriptstyle\bf v$}}}\in{\cal G}(DG)}\|{\mathchoice{\mbox{\boldmath$\displaystyle\bf v$}}{\mbox{\boldmath$\textstyle\bf v$}}{\mbox{\boldmath$\scriptstyle\bf v$}}{\mbox{\boldmath$\scriptscriptstyle\bf v$}}}\|_{1}\leq\|A\|_{\infty}^{st}\cdot(r\|DG\|_{\infty})^{r}.$$

$$\mathcal{L}(i,j)=(i,j)\times [-g(E):g(E)]\times [-g(E):g(E)]\times \Sigma (E).$$

$$L_{\max }\le g(E{)}^2\cdot \left|\Sigma \right|\le {\left(t^2(2r\Delta {)}^r\right)}^2\cdot {\left(1+4\Delta \cdot \left(t^2(2r\Delta {)}^r\right)\right)}^r\le {\left(t^2(2r\Delta {)}^r\right)}^{O(r)}.$$

$$\|\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}}-{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}^{*}\|_{\infty}\leq nt\cdot\max\left\{\|{\mathchoice{\mbox{\boldmath$\displaystyle\bf g$}}{\mbox{\boldmath$\textstyle\bf g$}}{\mbox{\boldmath$\scriptstyle\bf g$}}{\mbox{\boldmath$\scriptscriptstyle\bf g$}}}\|_{\infty}\mid{\mathchoice{\mbox{\boldmath$\displaystyle\bf g$}}{\mbox{\boldmath$\textstyle\bf g$}}{\mbox{\boldmath$\scriptstyle\bf g$}}{\mbox{\boldmath$\scriptscriptstyle\bf g$}}}\in{\cal G}\big{(}E^{(n)}\big{)}\right\}.$$

$$\min\big{\{}f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\mid E^{(n)}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}={\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}},{\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}\big{\}}=\min\big{\{}f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\mid E^{(n)}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}={\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}},\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}}\big{\}},$$

$$\displaystyle\min\leavevmode\nobreak\$$

$$\displaystyle\sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}\in 2^{U}}x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf f$}}{\mbox{\boldmath$\textstyle\bf f$}}{\mbox{\boldmath$\scriptstyle\bf f$}}{\mbox{\boldmath$\scriptscriptstyle\bf f$}}}}^{\tau}\leq n\qquad\mbox{for all }\tau\in[T];$$

$$\displaystyle\sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}\in{\cal T}_{c}}\ \sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf\alpha$}}{\mbox{\boldmath$\textstyle\bf\alpha$}}{\mbox{\boldmath$\scriptstyle\bf\alpha$}}{\mbox{\boldmath$\scriptscriptstyle\bf\alpha$}}}=({\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}^{\prime},f)\in{\cal T}_{s}}\delta(f,e^{\prime}_{i})x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf\alpha$}}{\mbox{\boldmath$\textstyle\bf\alpha$}}{\mbox{\boldmath$\scriptstyle\bf\alpha$}}{\mbox{\boldmath$\scriptscriptstyle\bf\alpha$}}}}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}}$$

$$\min b\cdot\left(\sum_{i=1}^{k}\ \sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}\in{\cal T}_{c}}\ \sum_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf\alpha$}}{\mbox{\boldmath$\textstyle\bf\alpha$}}{\mbox{\boldmath$\scriptstyle\bf\alpha$}}{\mbox{\boldmath$\scriptscriptstyle\bf\alpha$}}}=({\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}^{\prime},f)\in{\cal T}_{s}}\delta(f,e^{\prime}_{i})x_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf\alpha$}}{\mbox{\boldmath$\textstyle\bf\alpha$}}{\mbox{\boldmath$\scriptstyle\bf\alpha$}}{\mbox{\boldmath$\scriptscriptstyle\bf\alpha$}}}}^{{\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}}\right)\enspace.$$

$$\min \sum _{i=1}^T\sum _{j=1}^{|C|!}{\sigma }^i(i,j)x_j^i\text{leavevmode}\text{mbox}subjectto\text{leavevmode}\sum _{j=1}^{|C|!}{x}_j^i=n_i\text{leavevmode}\text{mbox}forall\text{leavevmode}i\in [T].$$

$$\sum _{i=1}^T\sum _{j=1}^{|C|!}{s}_j(c)x_j^i$$

$$\sum _{i=1}^T\sum _{j=1}^{|C|!}{\alpha }_j(c,c^{\prime })x_j^i$$

$${\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}=b\Leftrightarrow{\mathchoice{\mbox{\boldmath$\displaystyle\bf e$}}{\mbox{\boldmath$\textstyle\bf e$}}{\mbox{\boldmath$\scriptstyle\bf e$}}{\mbox{\boldmath$\scriptscriptstyle\bf e$}}}{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}-e_{j}^{i}l_{j}^{i}=b-e_{j}^{i}l_{j}^{i}\enspace.$$

$$\min$$

$$\displaystyle DC{\mathchoice{\mbox{\boldmath$\displaystyle\bf y$}}{\mbox{\boldmath$\textstyle\bf y$}}{\mbox{\boldmath$\scriptstyle\bf y$}}{\mbox{\boldmath$\scriptscriptstyle\bf y$}}}$$

$$\displaystyle\mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath$\displaystyle\bf y$}}{\mbox{\boldmath$\textstyle\bf y$}}{\mbox{\boldmath$\scriptstyle\bf y$}}{\mbox{\boldmath$\scriptscriptstyle\bf y$}}}^{i}$$

$$\displaystyle y_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf c$}}{\mbox{\boldmath$\textstyle\bf c$}}{\mbox{\boldmath$\scriptstyle\bf c$}}{\mbox{\boldmath$\scriptscriptstyle\bf c$}}}}^{i}$$

$$\displaystyle 0\leq y_{{\mathchoice{\mbox{\boldmath$\displaystyle\bf c$}}{\mbox{\boldmath$\textstyle\bf c$}}{\mbox{\boldmath$\scriptstyle\bf c$}}{\mbox{\boldmath$\scriptscriptstyle\bf c$}}}}^{i}$$

$$\begin{array}[]{ccccl|cccccl}D_{1}&D_{2}&\cdots&D_{k^{k}}&\leq d&\nobreak\leavevmode\hfil&D&D&\cdots&D&\leq d\\ \mathbf{1}^{\intercal}&0&\cdots&0&=b^{1}&&\mathbf{1}^{\intercal}&0&\cdots&0&=b^{1}\\ 0&\mathbf{1}^{\intercal}&\cdots&0&=b^{2}&&0&\mathbf{1}^{\intercal}&\cdots&0&=b^{2}\\ \vdots&\vdots&\ddots&\vdots&=\vdots&&\vdots&\vdots&\ddots&\vdots&=\vdots\\ 0&0&\cdots&\mathbf{1}^{\intercal}&=b^{k^{k}}&&0&0&\cdots&\mathbf{1}^{\intercal}&=b^{k^{k}},\\ \end{array}$$

$$F=\left[\begin{array}[]{cccc}D_{1}&D_{2}&\cdots&D_{\tau}\\ \mathbf{1}^{\intercal}&0&\cdots&0\\ 0&\mathbf{1}^{\intercal}&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&\mathbf{1}^{\intercal}\\ \end{array}\right]\enspace.$$

$$\min\big{\{}f({\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}})\mid F{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\diamondsuit{\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}},{\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}\leq{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}\big{\}}\enspace.$$

$$E^{(n)}=\left[\begin{array}[]{cccc}D&D&\cdots&D\\ \mathbf{1}^{\intercal}&0&\cdots&0\\ 0&\mathbf{1}^{\intercal}&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&\mathbf{1}^{\intercal}\\ \end{array}\right]\enspace.$$

$$\min\big{\{}\hat{f}(\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}})\mid E^{(n)}\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}}\diamondsuit{\mathchoice{\mbox{\boldmath$\displaystyle\bf b$}}{\mbox{\boldmath$\textstyle\bf b$}}{\mbox{\boldmath$\scriptstyle\bf b$}}{\mbox{\boldmath$\scriptscriptstyle\bf b$}}},\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf l$}}{\mbox{\boldmath$\textstyle\bf l$}}{\mbox{\boldmath$\scriptstyle\bf l$}}{\mbox{\boldmath$\scriptscriptstyle\bf l$}}}}\leq\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}}\leq\hat{{\mathchoice{\mbox{\boldmath$\displaystyle\bf u$}}{\mbox{\boldmath$\textstyle\bf u$}}{\mbox{\boldmath$\scriptstyle\bf u$}}{\mbox{\boldmath$\scriptscriptstyle\bf u$}}}}\big{\}}$$

Full text

Inspired by the work of Hemmecke, Onn and Romanchuk [Math. Prog. 2013], we develop a single-exponential algorithm for so-called combinatorial $n$-fold integer programs, which are remarkably similar to prior ILP formulations for various problems, but unlike them, also allow variable dimension. We then apply our algorithm to a few representative problems like Closest String, Swap Bribery, Weighted Set Multicover, and obtain exponential speedups in the dependence on the respective parameters, the input size, or both.

Unlike Lenstra’s algorithm, which is essentially a bounded search tree algorithm, our result uses the technique of augmenting steps. At its heart is a deep result stating that in combinatorial $n$-fold IPs, existence of an augmenting step implies existence of a “local” augmenting step, which can be found using dynamic programming. Our results provide an important insight into many problems by showing that they exhibit this phenomenon, and highlights the importance of augmentation techniques.

1 Introduction

The Integer Linear Programming (ILP) problem is fundamental as it models many combinatorial optimization problems. Since it is NP-complete, we naturally ask about the complexity of special cases. A fundamental algorithm by Lenstra from 1983 shows that ILPs can be solved in polynomial time when their number of variables (the dimension) $d$ is fixed [Len83]; that algorithm is thus a natural tool to prove that the complexity of some special cases of other NP-hard problems is also polynomial.

A systematic way to study the complexity of “special cases” of NP-hard problems was developed in the past 25 years in the field of parameterized complexity. There, the problem input is augmented by some integer parameter $k$, and one then measures the problem complexity in terms of both the instance size $n$ as well as $k$. Of central importance are algorithms with run times of the form f(k)n^{O\mathopen{}\mathclose{{}\left(1}\right)} for some computable function $f$, which are called fixed-parameter algorithms; the key idea is that the degree of the polynomial does not grow with $k$. For background on parameterized complexity, we refer to the monograph [CFK*+*15].

Kannan’s improvement [Kan87] of Lenstra’s algorithm runs in time d^{O\mathopen{}\mathclose{{}\left(d}\right)}n, which is thus a fixed-parameter algorithm for parameter $d$. Gramm et al. [GNR03] pioneered the application of Lenstra’s and Kannan’s algorithm in parameterized complexity: they modeled Closest String with $k$ input strings as an ILP of dimension k^{O\mathopen{}\mathclose{{}\left(k}\right)}, and thereby concluded with the first fixed-parameter algorithm for Closest String. This success led Niedermeier [Nie04] to propose in his book:

[…] It remains to investigate further examples besides Closest String where the described ILP approach turns out to be applicable. More generally, it would be interesting to discover more connections between fixed-parameter algorithms and (integer) linear programming.

Since then, many more applications of Lenstra’s and Kannan’s algorithm for parameterized problems have been proposed. However, essentially all of them [BFN*+*15, DS12, FLM*+*08, HR15, Lam12, MW15] share a common trait with the algorithm for Closest String: they have a doubly-exponential run time dependence on the parameter. Moreover, it is difficult to find ILP formulations with small dimension for problems whose input contains many objects with varying cost functions, such as in Swap Bribery [BCF*+*14, Challenge #2].

1.1 Our contributions

We show that a certain form of ILP, which is closely related to the previously used formulations for Closest String and other problems, can be solved in single-exponential time and in variable dimension. For example, Gramm et al.’s [GNR03] algorithm for Closest String runs in time 2^{2^{O\mathopen{}\mathclose{{}\left(k\log k}\right)}}O(\log L) for $k$ strings of length $L$ and has not been improved since 2003, while our algorithm runs in time k^{O\mathopen{}\mathclose{{}\left(k^{2}}\right)}O\mathopen{}\mathclose{{}\left(\log L}\right). Moreover, our algorithm has a strong combinatorial flavor and is based on different notions than are typically encountered in parameterized complexity, most importantly augmenting steps.

As an example of our form of ILP, consider the following ILP formulation of the Closest String problem. We are given $k$ strings $s_{1},\dots,s_{k}$ of length $L$ that come (after some preprocessing) from alphabet $[k]:=\{1,\ldots,k\}$, and an integer $d$. The goal is to find a string $y\in[k]^{L}$ such that, for each $s_{i}$, the Hamming distance $d_{H}(y,s_{i})$ is at most $d$, if such $y$ exists. For $i\in[L]$, $(s_{1}[i],\dots,s_{k}[i])$ is the $i$-th column of the input. Clearly there are at most $k^{k}$ different column types in the input, and we can represent the input succinctly with multiplicities b^{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}} of each column type {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\in[k]^{k}. Moreover, there are $k$ choices for the output string $y$ in each column. Thus, we can encode the solution by, for each column type {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\in[k]^{k} and each output character $e\in[k]$, describing how many solution columns are of type ({\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}},e). This is the basic idea behind the formulation of Gramm et al. [GNR03], as depicted on the left:

[TABLE]

Let $(1\leavevmode\nobreak\ \cdots\leavevmode\nobreak\ 1)=\mathbf{1}^{\intercal}$ be a row vector of all ones. Then we can view the above as

[TABLE]

where $D=(D_{1}\leavevmode\nobreak\ D_{2}\leavevmode\nobreak\ \dots\leavevmode\nobreak\ D_{k^{k}})$. The formulation on the right is clearly related to the one on the left, but contains “dummy” variables which are always zero. This makes it seem unnatural at first, but notice that it has the nice form

[TABLE]

Here, $r,s,t,n\in\mathbb{N}$, {\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\in\mathbb{Z}^{nt}, {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}\in\mathbb{Z}^{r+ns} and $f:\mathbb{Z}^{nt}\to\mathbb{Z}$ is a separable convex function (i.e., f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})=\sum_{i=1}^{n}\sum_{j=1}^{t}f^{i}_{j}(x^{i}_{j}) with every $f^{i}_{j}:\mathbb{Z}\to\mathbb{Z}$ univariate convex), $E^{(n)}$ is an $(r+ns)\times nt$-matrix, $D\in\mathbb{Z}^{r\times t}$ is an $r\times t$-matrix and $A\in\mathbb{Z}^{s\times t}$ is an $s\times t$-matrix. We call $E^{(n)}$ the $n$-fold product of $E=\left(\begin{smallmatrix}D\\ A\end{smallmatrix}\right)$. By L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle we denote the length of the binary encoding of the vectors {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}} and the objective function $f$, where \langle f\rangle=\langle\max_{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}:{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}|f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})|\rangle is the encoding length of the maximum absolute value attained by $f$ over the feasible region. This problem (1) is known as $n$-fold integer programming (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. Building on a dynamic program of Hemmecke, Onn and Romanchuk [HOR13] and a so-called proximity technique of Hemmecke, Köppe and Weismantel [HKW14], Knop and Koutecký [KK17] prove that:

Proposition 1** ([KK17, Thm. 7]).**

There is an algorithm that solves111Given an IP, we say that to solve it is to either (i) declare it infeasible or unbounded or (ii) find a minimizer of it. (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} encoded with L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle bits in time $\Delta^{O(trs+t^{2}s)}\cdot n^{3}L$, where $\Delta=1+\max\{\|D\|_{\infty},\|A\|_{\infty}\}$.

However, since the ILP on the right for Closest String satisfies $t=k^{k}$, applying Proposition 1 gives no advantage over applying Lenstra to solve the Closest String problem.

We overcome this impediment by harnessing the special structure of the ILP for Closest String. Observe that its constraint matrix $A$ has the form

[TABLE]

Moreover, under suitable assumptions on the objective function $f$, we call any such IP a combinatorial $n$-fold IP:

Definition 2** (Combinatorial $n$-fold IP).**

Let $A=(1\leavevmode\nobreak\ \cdots\leavevmode\nobreak\ 1)\in\mathbb{Z}^{1\times t}$, let $D\in\mathbb{Z}^{r\times t}$ be a matrix, and let $E=\left(\begin{smallmatrix}D\\ A\end{smallmatrix}\right)$. Let $f\colon\mathbb{Z}^{nt}\to\mathbb{Z}$ be a separable convex function represented by an evaluation oracle. A combinatorial $n$-fold IP is

[TABLE]

To use $f$ algorithmically, we also want $f$ to admit an efficient optimization oracle for the continuous relaxation of (9). This property we can often assume. Indeed, if for any $\alpha\in\mathbb{R}$ there is an efficient separation oracle for the level set \{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\mid f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})\leq\alpha\}, then the continuous optimum \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}} of \min\big{\{}f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})\mid E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}\big{\}} can be found in polynomial time using the ellipsoid method.

Our main result is a fast algorithm for combinatorial $n$-fold IPs, which is exponentially faster in $t$ than previous works for general $n$-fold IPs.

Theorem 3**.**

There is an algorithm that solves any combinatorial $n$-fold IP (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} of size L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle in time t^{O\mathopen{}\mathclose{{}\left(r}\right)}(\Delta r)^{O\mathopen{}\mathclose{{}\left(r^{2}}\right)}O\mathopen{}\mathclose{{}\left(n^{3}L}\right)+\mathsf{oo}, where $\Delta=1+\|D\|_{\infty}$ and $\mathsf{oo}$ is the time required for one call to an optimization oracle for the continuous relaxation of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}.

Observe that, when applicable, our algorithm is not only asymptotically faster than Lenstra’s, but works even if $n$ is variable (not parameter).

1.2 Succinctness

A common aspect shared by all of our applications is that bounding some parameter of the instance makes it preferable to view the instance in a succinct way (following the terminology of Faliszewski et al. [FHH06]; Onn [Onn14, OS15] calls these problems huge whereas Goemans and Rothvoß [GR14] call them high multiplicity). The standard way of viewing an instance is that the input is a collection of individual objects (bricks, matrix columns, voters, covering sets etc.). The succinct way of viewing an instance is by saying that identical objects are of the same type, giving a bound $T$ on the number of distinct types, and then presenting the input as numbers $n_{1},\dots,n_{T}$ such that $n_{i}$ is the number of objects of type $i$. Clearly, any standard instance can be converted to a succinct instance of roughly the same size (the number of objects is an upper bound on $T$), but the converse is not true as the numbers $n_{i}$ might be large. Also, it is sometimes not trivial (see Sect. 4.4) that the output can be represented succinctly; still, in all cases which we study it can.

In our applications we always state what are the types and what is the upper bound $T$ on the number of types; we assume some arbitrary enumeration of the types. We also assume that the input is presented succinctly and thus we do not include the time needed to read and convert a standard instance into a succinct instance in the runtime of our algorithms.

1.3 Applications

We apply Theorem 3 to several fundamental combinatorial optimization problems, for which we obtain exponential improvements in the parameter dependence, the input length, or both. For a summary of results, see Table 1; this list is not meant to be exhaustive. In fact, we believe that for any Lenstra-based result in the literature which only achieves double-exponential run times, there is a good chance that it can be sped up using our algorithm. The only significant obstacle seem to be large coefficients in the constraint matrix or an exponential number of “global” constraints.

Stringology.

A typical problem from stringology is to find a string $y$ satisfying certain distance properties with respect to $k$ strings $s_{1},\dots,s_{k}$. All previous fixed-parameter algorithms for such problems we are aware of for parameter $k$ rely on Lenstra’s algorithm, or their complexity status was open (e.g., the complexity of Optimal Consensus [ALN*+*11] was unknown for all $k\geq 4$). Interestingly, Boucher and Wilkie [BW10] show the counterintuitive fact that Closest String is easier to solve when $k$ is large, which makes the parameterization by $k$ even more significant. Finding an algorithm with run time only single-exponential in $k$ was a repeatedly posed problem, e.g. by Bulteau et al. [BHKN14, Challenge #1] and Avila et al. [AGST06, Problem 7.1]. By applying our result, we close this gap for a wide range of problems.

Theorem 4**.**

The problems

• •

Closest String*, Farthest String, Distinguishing String Selection, Neighbor String, Closest String with Wildcards, Closest to Most Strings, $c$-HRC and Optimal Consensus are solvable in time $k^{O(k^{2})}O(\log L)$, and,*

• •

$d$-Mismatch* is solvable in time $k^{O(k^{2})}O(L^{2}\log L)$,*

for inputs consisting of $k$ strings of length $L$ succinctly encoded by multiplicities of identical columns.

Computational social choice.

A typical problem in computational social choice takes as input an election consisting of a set $V$ of voters and a set $C$ of candidates which are ranked by the voters; the objective is to manipulate the election in certain ways to let a desired candidate win the election under some voting rule $\mathcal{R}$. This setup leads to a class of bribery problems, a prominent example of which is $\mathcal{R}$-Swap Bribery where manipulation is by swaps of candidates which are consecutive in voters’ preference orders. For a long time, the only known algorithms minimizing the number of swaps required run times which were doubly-exponential in $|C|$; improving those run times was posed as a challenge [BCF*+*14, Challenge #1]. Recently, Knop et al. [KKM17b] solved the challenge using Proposition 1. However, Knop et al.’s result has a cubic dependence $O(|V|^{3})$ on the number of voters, and the dependence on the number of candidates is still quite large, namely $|C|^{O(|C|^{6})}$.

We improve their result to logarithmic dependence on $|V|$, and smaller dependence on $|C|$. By $T$ we denote the number of voter types, where two voters are of the same type if they have the same preferences over the candidates and the same cost function for bribery; clearly $T\leq|V|$.

Theorem 5**.**

${\cal R}$-Swap Bribery can be solved in time

• •

$|C|^{O(|C|^{2})}O(T^{3}(\log|V|+\log\sigma_{\max}))$* for ${\cal R}$ any natural scoring protocol, and,*

• •

$|C|^{O(|C|^{4})}O(T^{3}(\log|V|+\log\sigma_{\max}))$* for ${\cal R}$ any C1 rule,*

where $T$ is the number of voter types and $\sigma_{\max}$ is the maximum cost of a swap.

Connections between stringology and computational social choice.

Challenge #3 of Bulteau et al. [BHKN14] asks for connections between problems in stringology and computational social choice. We demonstrate that in both fields combinatorial $n$-fold IP is an important tool. An important feature of both Bribery-like problems and Closest String-like problems is that permuting voters or characters does not have any effect. This fits well the $n$-fold IP format, which does not allow any interaction between bricks. It seems that this feature is important, as when it is taken away, such as in Closest Substring, the problem becomes $\mathsf{W}[1]$-hard [Mar08], even for parameter $d+k$.

Another common feature is that both types of problems naturally admit ILP formulations for succinct variants of the problems, as mentioned above. Moreover, it was precisely this fact that made all previous algorithms doubly-exponential—the natural succinct formulation has exponentially many (in the parameter) variables and thus applying Lenstra’s algorithm leads to a doubly-exponential runtime.

Weighted set multicover.

Bredereck et al. [BFN*+*15] define the Weighted Set Multicover (WSM) problem, which is a significant generalization of the classical Set Cover problem. Their motivation to study WSM was that it captures several problems from computational social choice and optimization problems on graphs implicit to previous works [FGK*+*17, GLO13, Lam12]. Bredereck et al. [BFN*+*15] design an algorithm for WSM that runs in time $2^{2^{O(k\log k)}}O(n)$, using Lenstra’s algorithm.

Again, our result yields an exponential improvement of that by Bredereck et al. [BCF*+*14], both in the dependence on the parameter and the size of the instance:

Theorem 6**.**

There is an algorithm that solves Weighted Set Multicover in time $k^{O(k^{2})}O(\log n+\log w_{\max})$ for succinctly represented instances of $n$ sets over a universe of size $k$, where $w_{\max}$ is the maximum weight of any set.

Huge $n$-fold IP.

Onn [Onn14] introduces a high-multiplicity version of the standard $n$-fold IP problem (1), where the number of bricks is now given in binary. It thus closely relates to the Cutting Stock problem, the high-multiplicity version of Bin Packing where the number of items for given size is given in binary; the complexity of Cutting Stock for constantly many item sizes was a long-standing open problem that was recently shown to be polynomial-time solvable by Goemans and Rothvoß [GR14] and by Jansen and Klein [JK17]. Previously, Huge $n$-fold IP was shown to be fixed-parameter tractable when $D=I$ and $A$ is totally unimodular; using our result, we show that it is also fixed-parameter tractable when $D$ and $A$ are arbitrary, but the size of variable domains is bounded by a parameter.

1.4 Comparison with Lenstra’s algorithm

The basic idea behind Lenstra’s algorithm is the following. Given a system A{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}} it is possible to compute its volume and determine that it is either too large not to contain an integer point, or too small not to be flat in some direction. In the first case we are done; in the second case we can take $d$ slices of dimension $d-1$ and recurse into them, achieving a $d^{O(d)}n^{O(1)}$ runtime. Note that we only decide feasibility; optimization can be then done by binary search. On the other hand, the basic idea behind our algorithm is the following. We only focus on optimizing and later show that testing feasibility reduces to it. Starting from some feasible solution, the crucial observation is that if there is a step improving the objective, there is one which does not modify many variables, and can be found quickly by dynamic programming. Moreover, if the current solution is far from the optimum, then it is possible to make a long step, and polynomially many long steps will reach the optimum.

More concretely, consider the run of these two algorithms on an instance of Closest String consisting in $k$ strings each of length $L$. Lenstra’s algorithm essentially either determines that the bounds are loose enough that there must exist a solution, or (oversimplifying) determines that there is a column type {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\in[k]^{k} and a character $e\in[k]$ such that there are at most $k^{k}$ consecutive choices for how many times the solution contains character $e$ at a column of type $\textstyle\bf f$. Then, we recurse, obtaining a $2^{2^{O(k\log k)}}O(\log L)$-time algorithm. On the other hand, our algorithm views the problem as an optimization problem, so we think of starting with a string of all blanks which is trivially at distance [math] from any string, and the goal is to fill in all blanks such that the result is still in distance at most $d$ from the input strings. An augmenting step is a set of character swaps that decreases the number of blanks. The crucial observation is that if an augmenting step exists, then there is also one only changing few characters, and it can be found in time $k^{O(k^{2})}O(\log L)$. Thus (omitting details), we can iteratively find augmenting steps until we reach the optimum.

Related work.

Our main inspiration are augmentation methods based on Graver bases, especially a fixed-parameter algorithm for $n$-fold IP of Hemmecke, Onn and Romanchuk [HOR13]. Our result improves the runtime of their algorithm for a special case. All the following related work is orthogonal to ours in either the achieved result, or the parameters used for it.

In fixed dimension, Lenstra’s algorithm [Len83] was generalized for arbitrary convex sets and quasiconvex objectives by Khachiyan and Porkolab [KP00]. The currently fastest algorithm of this kind is due to Dadush et al. [DPV11]. The first notable fixed-parameter algorithm for a non-convex objective is due to Lokshtanov [Lok15], who shows that optimizing a quadratic function over the integers of a polytope is fixed-parameter tractable if all coefficients are small. Ganian and Ordyniak [GO16] and Ganian et al. [GOR17] study the complexity of ILP with respect to structural parameters such as treewidth and treedepth, and introduce a new parameter called torso-width.

Besides fixed-parameter tractability, there is interest in the (non)existence of kernels of ILPs, which formalize the (im)possibility of various preprocessing procedures. Jansen and Kratsch [JK15] show that ILPs containing parts with simultaneously bounded treewidth and bounded domains are amenable to kernelization, unlike ILPs containing totally unimodular parts. Kratsch [Kra16] studies the kernelizability of sparse ILPs with small coefficients.

2 Preliminaries

For positive integers $m,n$ we set $[m\,:\,n]=\{m,\ldots,n\}$ and $[n]=[1\,:\,n]$. For a graph $G$ we denote by $V(G)$ its set of vertices. We write vectors in boldface (e.g., {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf y}}{\mbox{\boldmath\textstyle\bf y}}{\mbox{\boldmath\scriptstyle\bf y}}{\mbox{\boldmath\scriptscriptstyle\bf y}}}) and their entries in normal font (e.g., the $i$-th entry of $\textstyle\bf x$ is $x_{i}$). If $A$ is a matrix, $A_{r}$ denotes its $r$-th column. For a matrix $A\in\mathbb{Z}^{m\times n}$, vectors {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}\in\mathbb{Z}^{m}, {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}\in\mathbb{Z}^{n} and a function $f:\mathbb{Z}^{n}\to\mathbb{Z}$, let (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} be the problem

[TABLE]

We say that $\textstyle\bf x$ is feasible for (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} if A{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}} and {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}.

Graver bases and augmentation.

Let us now introduce Graver bases and discuss how they can be used for optimization. We also recall $n$-fold IPs; for background, we refer to the books of Onn [Onn10] and De Loera et al. [DHK13].

Let {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf y}}{\mbox{\boldmath\textstyle\bf y}}{\mbox{\boldmath\scriptstyle\bf y}}{\mbox{\boldmath\scriptscriptstyle\bf y}}} be $n$-dimensional integer vectors. We call {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf y}}{\mbox{\boldmath\textstyle\bf y}}{\mbox{\boldmath\scriptstyle\bf y}}{\mbox{\boldmath\scriptscriptstyle\bf y}}} sign-compatible if they lie in the same orthant, that is, for each $i\in[n]$ the sign of $x_{i}$ and $y_{i}$ is the same. We call \sum_{i}{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{i} a sign-compatible sum if all {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{i} are pair-wise sign-compatible. Moreover, we write {\mathchoice{\mbox{\boldmath\displaystyle\bf y}}{\mbox{\boldmath\textstyle\bf y}}{\mbox{\boldmath\scriptstyle\bf y}}{\mbox{\boldmath\scriptscriptstyle\bf y}}}\sqsubseteq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}} if $\textstyle\bf x$ and $\textstyle\bf y$ are sign-compatible and $|y_{i}|\leq|x_{i}|$ for each $i\in[n]$, and write {\mathchoice{\mbox{\boldmath\displaystyle\bf y}}{\mbox{\boldmath\textstyle\bf y}}{\mbox{\boldmath\scriptstyle\bf y}}{\mbox{\boldmath\scriptscriptstyle\bf y}}}\sqsubset{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}} if at least one of the inequalities is strict. Clearly, $\sqsubseteq$ imposes a partial order called “conformal order” on $n$-dimensional vectors. For an integer matrix $A\in\mathbb{Z}^{m\times n}$, its Graver basis ${\cal G}(A)$ is the set of $\sqsubseteq$-minimal non-zero elements of the lattice of $A$, \ker_{\mathbb{Z}}(A)=\{{\mathchoice{\mbox{\boldmath\displaystyle\bf z}}{\mbox{\boldmath\textstyle\bf z}}{\mbox{\boldmath\scriptstyle\bf z}}{\mbox{\boldmath\scriptscriptstyle\bf z}}}\in\mathbb{Z}^{n}\mid A{\mathchoice{\mbox{\boldmath\displaystyle\bf z}}{\mbox{\boldmath\textstyle\bf z}}{\mbox{\boldmath\scriptstyle\bf z}}{\mbox{\boldmath\scriptscriptstyle\bf z}}}=\mathbf{0}\}. An important property of ${\cal G}(A)$ is the following.

Proposition 7** ([Onn10, Lemma 3.2]).**

Every integer vector {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\neq\mathbf{0} with A{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}=\mathbf{0} is a sign-compatible sum {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}=\sum_{i}{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{i} of Graver basis elements {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{i}\in{\cal G}(A), with some elements possibly appearing with repetitions.

Let $\textstyle\bf x$ be a feasible solution to (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. We call $\textstyle\bf g$ a feasible step if {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}} is feasible for (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. Further, call a feasible step $\textstyle\bf g$ augmenting if f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}})<f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}). An augmenting step $\textstyle\bf g$ and a step length $\alpha\in\mathbb{Z}$ form an $\textstyle\bf x$-feasible step pair with respect to a feasible solution $\textstyle\bf x$ if {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}. An augmenting step $\textstyle\bf g$ and a step length $\alpha\in\mathbb{Z}$ form a Graver-best step if f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}})\leq f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha^{\prime}\tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}) for all $\textstyle\bf x$-feasible step pairs (\tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}},\alpha^{\prime})\in{\cal G}(A)\times\mathbb{Z}.

The Graver-best augmentation procedure for (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} with given feasible solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0} works as follows:

If there is no Graver-best step for {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0}, return it as optimal. 2. 2.

If a Graver-best step (\alpha,{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}) for {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0} exists, set {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0}:={\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}} and go to 1.

Proposition 8** ([DHK13, implicit in Theorem 3.4.1]).**

Given a feasible solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0} for (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} where $f$ is separable convex, the Graver-best augmentation procedure finds an optimum of (IP)_{A,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} in at most $2n-2\log M$ steps, where M=f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0})-f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{*}) and {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{*} is any minimizer of $f$.

$n$-fold IP.

The structure of $E^{(n)}$ (in problem (1)) allows us to divide the $nt$ variables of $\textstyle\bf x$ into $n$ bricks of size $t$. We use subscripts to index within a brick and superscripts to denote the index of the brick, i.e., $x_{j}^{i}$ is the $j$-th variable of the $i$-th brick with $j\in[t]$ and $i\in[n]$.

3 Combinatorial $\boldsymbol{n}$-fold IPs

This section is dedicated to proving Theorem 3. We fix an instance of combinatorial $n$-fold IP, that is, a tuple (n,D,{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f).

3.1 Graver complexity of combinatorial $n$-fold IP

The key property of the $n$-fold product $E^{(n)}$ is that, for any $n\in\mathbb{N}$, the number of nonzero bricks of any {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} is bounded by some constant $g(E)$ called the Graver complexity of $E$. A proof is given for example by Onn [Onn10, Lemma 4.3]; it goes roughly as follows. Consider any {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} and take its restriction to its nonzero bricks \bar{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}. By Proposition 7, each brick \bar{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}^{j} can be decomposed into elements from ${\cal G}(A)$, giving a vector $\textstyle\bf h$ whose bricks are elements of ${\cal G}(A)$. Then, consider a compact representation $\textstyle\bf v$ of $\textstyle\bf h$ by counting how many times each element from ${\cal G}(A)$ appears. Since {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} and $\textstyle\bf h$ is a decomposition of its nonzero bricks, we have that \sum_{j}D{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{j}=0. Let $G$ be a matrix with the elements of ${\cal G}(A)$ as columns. It is not difficult to show that {\mathchoice{\mbox{\boldmath\displaystyle\bf v}}{\mbox{\boldmath\textstyle\bf v}}{\mbox{\boldmath\scriptstyle\bf v}}{\mbox{\boldmath\scriptscriptstyle\bf v}}}\in{\cal G}(DG). Since \|{\mathchoice{\mbox{\boldmath\displaystyle\bf v}}{\mbox{\boldmath\textstyle\bf v}}{\mbox{\boldmath\scriptstyle\bf v}}{\mbox{\boldmath\scriptscriptstyle\bf v}}}\|_{1} is an upper bound on the number of bricks of $\textstyle\bf h$ and thus of nonzero bricks of $\textstyle\bf g$ and clearly does not depend on $n$, g(E)=\max_{v\in{\cal G}(DG)}\|{\mathchoice{\mbox{\boldmath\displaystyle\bf v}}{\mbox{\boldmath\textstyle\bf v}}{\mbox{\boldmath\scriptstyle\bf v}}{\mbox{\boldmath\scriptscriptstyle\bf v}}}\|_{1} is finite.

Let us make precise two observations from this proof.

Lemma 9** ([HOW11, Lemma 3.1], [Onn10, implicit in proof of Lemma 4.3]).**

Let ({\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{1},\dots,{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{n})\in{\cal G}\big{(}E^{(n)}\big{)}. Then for $i=1,\ldots,n$ there exist vectors {\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{i,1},\dots,{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{i,n_{i}}\in{\cal G}(A) such that {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{i}=\sum_{k=1}^{n_{i}}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{i,k}, and $\sum_{i=1}^{n}n_{i}\leq g(E)$.

Lemma 10** ([HOW11, Lemma 6.1], [Onn10, implicit in proof of Lemma 4.3]).**

Let $D\in\mathbb{Z}^{r\times t}$, $A\in\mathbb{Z}^{s\times t}$ and let $G\in\mathbb{Z}^{t\times p}$ be the matrix whose columns are the elements of ${\cal G}(A)$. Then $|{\cal G}(A)|\leq\|A\|_{\infty}^{st}$, and for $E=\left(\begin{smallmatrix}D\\ A\end{smallmatrix}\right)$ it holds

[TABLE]

Notice that this bound on $g(E)$ is exponential in $t$. Our goal now is to improve the bound on $g(E)$ in terms of $t$, exploiting the simplicity of the matrix $A$ in combinatorial $n$-fold IPs.

To see this, we will need to understand the structure of ${\cal G}(\mathbf{1}^{\intercal})$:

Lemma 11**.**

It holds that

• •

{\cal G}(\mathbf{1}^{\intercal})=\{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\mid{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\mbox{ has one 1$and one$-1$and$0 otherwise}\}\subseteq\mathbb{Z}^{t},

• •

$p=|{\cal G}(\mathbf{1}^{\intercal})|=t(t-1)$,

• •

and \|{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\|_{1}=2 for all {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}(\mathbf{1}^{\intercal}).

Proof.

Observe that the claimed set of vectors is clearly $\sqsubseteq$-minimal in $\ker_{\mathbb{Z}}(\mathbf{1}^{\intercal})$. We are left with proving there is no other non-zero $\sqsubseteq$-minimal vector in $\ker_{\mathbb{Z}}(\mathbf{1}^{\intercal})$. For contradiction assume there is such a vector $\textstyle\bf h$. Since it is non-zero, it must have a positive entry $h_{i}$. On the other hand, since \mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}, it must also have a negative entry $h_{j}$. But then $\textstyle\bf g$ with $g_{i}=1$, $g_{j}=-1$ and $g_{k}=0$ for all $k\not\in\{i,j\}$ is {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\sqsubset{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}, a contradiction. The rest follows. ∎

With this lemma in hand, we can prove that:

Lemma 12**.**

Let $D\in\mathbb{Z}^{r\times t}$, $E=\left(\begin{smallmatrix}D\\ \mathbf{1}^{\intercal}\end{smallmatrix}\right)$, and $\Delta=1+\|D\|_{\infty}$. Then, $g(E)\leq t^{2}(2r\Delta)^{r}$.

Proof.

We simply plug the correct values into the bound of Lemma 10. By Lemma 11, $p=t(t-1)\leq t^{2}$. Also, \|DG\|_{\infty}\leq\max_{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}(\mathbf{1}^{\intercal})}\left\{\|D\|_{\infty}\cdot\|{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\|_{1}\right\}\leq 2\Delta where the last inequality follows from \|{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\|_{1}=2 for all {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}(\mathbf{1}^{\intercal}), again by Lemma 11. ∎

3.2 Dynamic programming

Hemmecke, Onn and Romanchuk [HOR13] devise a clever dynamic programming algorithm to find augmenting steps for a feasible solution of an $n$-fold IP. Lemma 9 is key in their approach, as they continue by building a set $Z(E)$ of all sums of at most $g(E)$ elements of ${\cal G}(A)$ and then use it to construct the dynamic program. However, such a set $Z(E)$ would clearly be of size exponential in $t$—too large to achieve our single-exponential run times. In their dynamic program, layers correspond to partial sums of elements of ${\cal G}(A)$.

Our insight is to build a different dynamic program. In our dynamic program, we will exploit the simplicity of ${\cal G}(A)={\cal G}(\mathbf{1}^{\intercal})$ so that the layers will immediately correspond to the coordinates $h^{i}_{j}$ of an augmenting vector $\textstyle\bf h$. Additionally, we also differ in how we enforce feasibility with respect to the upper rows $(D\leavevmode\nobreak\ D\leavevmode\nobreak\ \cdots\leavevmode\nobreak\ D)$.

We now give the details of our approach. Let $\Sigma(E)=\prod_{j=1}^{r}\left[-2\Delta\cdot g(E)\,:\,2\Delta\cdot g(E)\right]$ be the signature set of $E$ whose elements are signatures. Essentially, we will use the signature set to keep track of partial sums of prefixes of the augmenting vector $\textstyle\bf h$ to ensure that it satisfies D{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}. Crucially, we notice that to ensure D{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}, it suffices to remember the partial sum of the prefixes of $\textstyle\bf h$ multiplied by $D$, thus shrinking them to dimension $r$. This is another insight which allows us to avoid the exponential dependence on $t$. Note that $\left|\Sigma(E)\right|\leq\left(1+4g(E)\Delta\right)^{r}$.

Definition 13** (Augmentation graph).**

Let $\textstyle\bf x$ be a feasible solution for (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} and let ${\alpha\in\mathbb{N}}$. Their augmentation graph DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) is a vertex-weighted directed layered graph with two distinguished vertices $S$ and $T$ called the source and the sink; and $nt$ layers ${\cal L}(1,1),\ldots,{\cal L}(n,t)$ structured according to the bricks, such that for all $i\in[n]$, $j\in[t]$,

[TABLE]

Thus, each vertex is a tuple \big{(}i,j,h^{i}_{j},\beta^{i}_{j},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}\big{)}, with the following meaning:

• •

$i\in[n]$ is the index of the brick,

• •

$j\in[t]$ is the position within the brick,

• •

$h^{i}_{j}\in[-g(E)\,:\,g(E)]$ is the value of the corresponding coordinate of a proposed augmenting vector $\textstyle\bf h$,

• •

$\beta^{i}_{j}\in[-g(E)\,:\,g(E)]$ is a brick prefix sum $\sum_{\ell=1}^{j}h^{i}_{j}$ of the proposed augmenting vector $\textstyle\bf h$, and,

• •

{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}\in\Sigma(E) is the signature, representing the prefix sum \sum_{k=1}^{i}D{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{k}+\sum_{\ell=1}^{j}D_{\ell}h^{i}_{\ell}.

A vertex \big{(}i,j,h^{i}_{j},\beta^{i}_{j},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}\big{)} has weight $f^{i}_{j}(\alpha h^{i}_{j}+x^{i}_{j})-f^{i}_{j}(x^{i}_{j})$.

Let S=\big{(}0,t,0,0,\mathbf{0}\big{)} and T=\big{(}n+1,1,0,0,\mathbf{0}\big{)}, where the last coordinate is an $r$-dimensional all-zero vector.

Edges to the first layer of a brick. Every vertex \big{(}i,t,h^{i}_{t},0,{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{t}\big{)} has edges to each vertex \big{(}i+1,1,h^{i+1}_{1},h^{i+1}_{1},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i+1}_{1}\big{)} for which $l^{i+1}_{1}\leq x^{i+1}_{1}+\alpha h^{i+1}_{1}\leq u^{i+1}_{1}$ and {\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i+1}_{1}={\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{t}+D_{1}h^{i+1}_{1}. Recall that $D_{j}$ is the $j$-th column of matrix $D$. We emphasize that there are no other outgoing edges from layer ${\cal L}(i,t)$ to layer ${\cal L}(i+1,1)$.

Edges within a brick. Every vertex \big{(}i,j,h^{i}_{j},\beta^{i}_{j},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}) with $j<t$ has edges to each vertex \big{(}i,j+1,h^{i}_{j+1},\beta^{i}_{j+1},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j+1}\big{)} for which

• •

$l^{i}_{j+1}\leq x^{i}_{j+1}+\alpha h^{i}_{j+1}\leq u^{i}_{j+1}$,

• •

$\beta^{i}_{j+1}=\beta^{i}_{j}+h^{i}_{j+1}$, with $\beta^{i}_{j+1}\in\left[-g(E)\colon g(E)\right]$, and,

• •

{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j+1}={\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}+D_{j+1}h^{i}_{j+1}.

See Fig. 1 for a scheme of the augmentation graph.

Note that by the bounds on $g(E)$ by Lemma 12, the number of vertices in each layer of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) is bounded by

[TABLE]

Let $P$ be an $S$–$T$ path in DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha). We define the $P$-augmentation vector {\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}\in\mathbb{Z}^{nt} by the $h^{i}_{j}$-coordinates of the vertices of $P$.

Let $\textstyle\bf x$ be a feasible solution of the combinatorial $n$-fold IP instance fixed at the beginning of Sect. 3. We say that $\textstyle\bf h$ is a solution of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) if there exists an $S$–$T$ path $P$ such that $\textstyle\bf h$ is the $P$-augmentation vector. The weight w({\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}) is then defined as the weight of the path $P$; note that w({\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}})=f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}})-f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}).

The following lemma relates solutions of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) to potential feasible steps in {\cal G}\big{(}E^{(n)}\big{)}.

Lemma 14**.**

Let {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\in\mathbb{Z}^{nt} be a feasible solution, $\alpha\in\mathbb{N}$, and let $\textstyle\bf h$ be a solution of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha). Then {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}} and E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}.

Proof.

To see that {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}, recall that there is no incoming edge to a vertex (i,j,h^{i}_{j},\beta^{i}_{j},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}) which would violate the bound $l^{i}_{j}\leq x^{i}_{j}+\alpha h^{i}_{j}\leq u^{i}_{j}$.

To see that E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}, first observe that by the definition of $\beta^{i}_{j}$, and the condition that only if $\beta^{i}_{t}=0$ there is an outgoing edge, we have that every brick {\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{i} satisfies \mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}^{i}=0. Second, by the definition of {\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j} and the edges incoming to $T$, we have that D{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}. Together, this implies E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}. ∎

Lemma 15**.**

Let {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\in\mathbb{Z}^{nt} be a feasible solution. Then every {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} with {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}} is a solution of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha).

Proof.

Let {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} satisfy {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}. We shall construct an $S$–$T$ path $P$ in DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) such that $\textstyle\bf g$ is the $P$-augmentation vector. We will describe which vertex is selected from each layer, and argue that this well defined. Then, by the definition of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha), it will be clear that the selected vertices are indeed connected by edges.

In layer ${\cal L}(1,1)$, we select vertex \big{(}1,1,g^{1}_{1},g^{1}_{1},D_{1}g^{1}_{1}\big{)}. In layer ${\cal L}(i,j)$, we select vertex \big{(}i,j,g^{i}_{j},\beta^{i}_{j},{\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}\big{)} with $\beta^{i}_{j}=\beta^{i}_{j-1}+g^{i}_{j}$ and {\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}={\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j-1}+D_{j}g^{i}_{j} if $j>1$, and $\beta^{i}_{j}=\beta^{i-1}_{t}+g^{i}_{j}$ and {\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j}={\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i-1}_{t}+g^{i-1}_{t} otherwise.

We shall argue that this is well defined, i.e., that all of the specified vertices actually exist. From Lemma 9 it follows that $\textstyle\bf g$ can be decomposed into $M\leq g(E)$ vectors \tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}^{1},\dots\tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}^{M}\in{\cal G}(\mathbf{1}^{\intercal}). By Lemma 11, \|\tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}}^{i}\|_{1}\leq 2 for every $i$, which implies that \|{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\|_{\infty}\leq g(E). Moreover, since $M\leq g(E)$, we also have that for every $i\in[n]$ and every $j\in[t]$, $|\sum_{\ell=1}^{j}g^{i}_{j}|\leq g(E)$ (i.e., the brick prefix sum is also bounded by $g(E)$ in absolute value). Thus, vertices with the appropriate $g^{i}_{j}$- and $\beta^{i}_{j}$- coordinates exist. Regarding the {\mathchoice{\mbox{\boldmath\displaystyle\bf\sigma}}{\mbox{\boldmath\textstyle\bf\sigma}}{\mbox{\boldmath\scriptstyle\bf\sigma}}{\mbox{\boldmath\scriptscriptstyle\bf\sigma}}}^{i}_{j} coordinate, we make a similar observation: for every $i\in[n]$ and $j\in[t]$, \big{(}\sum_{\hat{\ell}=1}^{i-1}D{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}^{\hat{\ell}}\big{)}+\big{(}\sum_{\ell=1}^{j}D_{\ell}g^{i}_{\ell}\big{)}\in\Sigma(E).

From the definition of the edges in DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha), and the fact that {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}, the selected vertices create a path. ∎

Lemma 16** (optimality certification).**

There is an algorithm that, given a feasible solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\in\mathbb{Z}^{nt} for (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} and $\alpha\in\mathbb{N}$, in time $ntL_{\max}^{2}\leq t^{O(r)}(\Delta r)^{O(r^{2})}n$ either finds a vector $\textstyle\bf h$ such that (i) E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}=\mathbf{0}, (ii) {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}} and (iii) f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}})<f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}), or decides that none exists.

Proof.

It follows from Lemma 14 that all solutions of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}) fulfill (i) and (ii). Observe that if we take $\textstyle\bf h$ to be a solution of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) with minimum weight, then either f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})=f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}) or f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})>f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}). Due to Lemma 15 the set of solutions of DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) contains all {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} with {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}. Thus, by Proposition 8, if f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})=f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}), no $\textstyle\bf g$ satisfying all three conditions (i), (ii), (iii) exist.

Our goal is then to find the lightest $S$–$T$ path in the graph DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha). However, since edges of negative weight will be present, we cannot use, e.g., Dijkstra’s algorithm. Still, it can be observed that DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha) is a directed acyclic graph, and moreover, finding the lightest path can be done in a layer-by-layer manner in time O(|V(DP({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}},\alpha))|\cdot L_{\max})=O(ntL_{\max}^{2}); cf. [HOR13, Lemma 3.4]. The claimed run time follows from the bound on the maximum size of a layer (10). ∎

3.3 Step lengths

We have shown how to find a feasible step $\textstyle\bf h$ for any given step length $\alpha\in\mathbb{N}$ such that f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}})\leq f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}) for any feasible step {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)}. Now, we will show that there are not too many step lengths that need to be considered in order to find a Graver-best step which, by Proposition 8, leads to a good bound on the total required number of steps. This is the case in particular if we have an instance whose feasible solutions are all contained in a box {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}} with \|{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}\|_{\infty}\leq N, as no steps of length $\alpha>N$ are feasible.

In the following, we use the proximity technique pioneered by Hochbaum and Shantikumar [HS90] in the case of totally unimodular matrices and extended to the setting of Graver bases by Hemmecke, Köppe and Weismantel [HKW14]. This technique allows to show that, provided some structure of the constraints (e.g., total unimodularity or bounded $\ell_{\infty}$-norm of its Graver elements), the continuous optimum is not too far from the integer optimum of the problem.

Proposition 17**.**

*([HKW14, Theorem 3.14])

Consider a combinatorial $n$-fold IP (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. Then for any optimal solution \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}} of the continuous relaxation of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} there is an optimal solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{*} of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} with*

[TABLE]

This allows us then to reduce the original instance to an equivalent instance contained in a small box, as hinted at above.

Lemma 18** (equivalent bounded instance).**

Let (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} be a combinatorial $n$-fold IP of size L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle. With one call to an optimization oracle of its continuous relaxation, one can construct \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}},\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}\in\mathbb{Z}^{nt} such that

[TABLE]

and \|\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}-\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}}\|_{\infty}\leq nt\cdot g(E).

Proof.

Returning to Proposition 17, observe that the quantity \max\left\{\|{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\|_{\infty}\mid{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)}\right\} is bounded by $g(E)$ (Lemma 11). Hence, we can set new lower and upper bounds \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}} and \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}} defined by $\hat{l}_{j}^{i}:=\max\left\{\lfloor\hat{x}_{j}^{i}\rfloor-ng(E),l_{j}^{i}\right\}$ and $\hat{u}_{j}^{i}:=\min\left\{\lceil\hat{x}_{j}^{i}\rceil+ng(E),u_{j}^{i}\right\}$, and Theorem 17 assures that the integer optimum also lies within the new bounds. ∎

3.4 Finishing the proof

Proof of Theorem 3..

We proceed in three steps.

Step 1: Bounding the feasible region. First, we use Lemma 18 to construct new lower and upper bounds \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}},\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}} satisfying \|\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}-\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}}\|_{\infty}\leq nt\cdot g(E) and preserving the optimal value of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. Thus, we shall replace $\textstyle\bf l$ and $\textstyle\bf u$ by \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}} and \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}} from now on and assume that \|{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\|_{\infty}\leq nt\cdot g(E).

Step 2: Optimization. Let us assume that we have an initial feasible solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0} of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. Given a step length $\alpha\in\mathbb{N}$, it takes time $t^{O(r)}(\Delta r)^{O(r^{2})}n$ by Lemma 16 to find a feasible step $\textstyle\bf h$ satisfying f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}})\leq f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}) for all {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)}. Recall that no {\mathchoice{\mbox{\boldmath\displaystyle\bf g}}{\mbox{\boldmath\textstyle\bf g}}{\mbox{\boldmath\scriptstyle\bf g}}{\mbox{\boldmath\scriptscriptstyle\bf g}}}\in{\cal G}\big{(}E^{(n)}\big{)} can be feasible for $\alpha>ng(E)$ by our bound on \|{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\|_{\infty}. Thus, applying Lemma 16 for all $\alpha\in[nt\cdot g(E)]$ and choosing a pair (\alpha,{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}) which minimizes f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}+\alpha{\mathchoice{\mbox{\boldmath\displaystyle\bf h}}{\mbox{\boldmath\textstyle\bf h}}{\mbox{\boldmath\scriptstyle\bf h}}{\mbox{\boldmath\scriptscriptstyle\bf h}}}) surely finds a Graver-best step in time $t^{O(r)}(\Delta r)^{O(r^{2})}n^{2}$. In order to reach the optimum, by Proposition 8 we need to make at most $(2nt-2)\cdot O(L)$ Graver-best steps, where L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},\mathbf{0},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle; this is because $O(L)$ is an upper bound on f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{0})-f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{*}) for some optimal solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{*} of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}. In total, we need time $t^{O(r)}(\Delta r)^{O(r^{2})}n^{3}L$.

Step 3: Feasibility. Now we are left with the task of finding a starting feasible solution {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}_{0} in the case when we do not have it. We follow the lines of Hemmecke et al. [HOR13, Lemma 3.8] and solve an auxiliary combinatorial $n$-fold IP given by the bimatrix $\bar{E}=\left(\begin{smallmatrix}\bar{D}\\ \bar{A}\end{smallmatrix}\right)$ with $\bar{D}:=(D\leavevmode\nobreak\ I_{r}\leavevmode\nobreak\ -I_{r}\leavevmode\nobreak\ \mathbf{0})$ and $\bar{A}:=(A\leavevmode\nobreak\ \mathbf{1}_{2r+1})=\mathbf{1}^{\intercal}\in\mathbb{Z}^{t+2r+1}$, where $I_{r}$ is the identity matrix of dimension $r$, $\mathbf{0}$ is a column vector of length $r$ and $\mathbf{1}_{2r+1}$ is the vector of all $1$s of length $2r+1$. The variables \bar{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}} of this auxiliary problem have a natural partition into $nt$ variables $\textstyle\bf x$ corresponding to the original problem fixed at the beginning of Sect. 3, and $n(2r+1)$ new auxiliary variables \tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}}. Keep the original lower and upper bounds on $\textstyle\bf x$ and introduce a lower bound [math] and upper bound $ntg(E)\Delta$ on each auxiliary variable. Finally, let the new objective be f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}})=\bar{{\mathchoice{\mbox{\boldmath\displaystyle\bf w}}{\mbox{\boldmath\textstyle\bf w}}{\mbox{\boldmath\scriptstyle\bf w}}{\mbox{\boldmath\scriptscriptstyle\bf w}}}}^{\intercal}\bar{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}}, i.e., the sum of the auxiliary variables. Observe that it is easy to construct an initial feasible solution by setting {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}} and computing \tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}} accordingly: \tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}} serve the role of slack variables, and the slack in any constraint is at most $nt^{2}g(E)\Delta$ by the fact that \|{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}\|_{\infty}\leq nt\cdot g(E) and $\Delta=1+\|D\|_{\infty}$.

Then, applying the optimization algorithm described in the beginning of the proof either finds a solution to the auxiliary problem with objective value [math], implying \tilde{{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}}=\mathbf{0}, and thus $\textstyle\bf x$ is feasible for the original problem, or no such solution exists, meaning that the original problem is infeasible. ∎

4 Applications

Let us repeat the main result of this paper:

Theorem 3 (repeated).

There is an algorithm that solves any combinatorial $n$-fold IP (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f} of size L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle in time t^{O\mathopen{}\mathclose{{}\left(r}\right)}(\Delta r)^{O\mathopen{}\mathclose{{}\left(r^{2}}\right)}O\mathopen{}\mathclose{{}\left(n^{3}L}\right)+\mathsf{oo}, where $\Delta=1+\|D\|_{\infty}$ and $\mathsf{oo}$ is the time required for one call to an optimization oracle for the continuous relaxation of (IP)_{E^{(n)},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f}.

In applications, it is practical to use combinatorial $n$-fold IP formulations which contain inequalities. Given an $n$-fold IP (in particular a combinatorial $n$-fold IP), we call the upper rows (D\leavevmode\nobreak\ D\leavevmode\nobreak\ \cdots\leavevmode\nobreak\ D){\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{0} globally uniform constraints, and the lower rows A{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{i}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{i}, for all $i\in[n]$, locally uniform constraints. So we first show that introducing inequalities into a combinatorial $n$-fold IP is possible. However in the case of globally uniform constraints, we need a slightly different approach than in a standard $n$-fold IP to keep the rigid format of a combinatorial $n$-fold IP.

Inequalities in locally uniform constraints.

We add $n$ variables $x_{t+1}^{i}$ for all $i\in[n]$ and we replace $D$ with $(D\leavevmode\nobreak\ \mathbf{0})$. For each row $i$ where we wish to enforce \mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{i}\leq b^{i}, we set the upper bound on $u_{t+1}^{i}=b^{i}$ and lower bound $l_{t+1}^{i}=0$. Similarly, for each row $i$ where we wish to enforce \mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{i}\geq b^{i}, we set a lower bound $l_{t+1}^{i}=-b^{i}$ and an upper bound $u_{t+1}^{i}=0$. For all remaining rows we set $l_{t+1}^{i}=u_{t+1}^{i}=0$, enforcing \mathbf{1}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{i}=b^{i}.

Inequalities in globally uniform constraints.

We replace $D$ with $(D\leavevmode\nobreak\ I_{r})$, where $I_{r}$ is the $r\times r$ identity matrix. Thus, we have introduced $r$ new variables $x_{t+j}^{i}$ with $i\in[n]$ and $j\in[r]$; however, we enforce them all to be [math] by setting $l_{t+j}=u_{t+j}^{i}=0$ for all $i\in[n]$ and $j\in[r]$. Next, we introduce an $(n+1)$-st brick, set $u_{j}^{n+1}=0$ for all $j\in[t]$ and set $b^{n+1}=\|D\|_{\infty}\cdot\|(b^{1},\dots,b^{n})\|_{1}$. Then, for each row $i\in[r]$ where we wish to enforce a “$\leq$” inequality, we set $l_{t+i}^{n+1}=0$ and u_{t+i}^{n+1}=\|{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}\|_{\infty}, and for each row $i\in[r]$ with a “$\geq$” inequality, we set l_{t+i}^{n+1}=-\|{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}\|_{\infty} and $u_{t+i}^{n+1}=0$. We let $l_{t+i}^{n+1}=u_{t+r+i}^{n+1}=0$ for equality. To enforce a strict inequality “$<$”, we proceed as for “$\leq$” and increase the corresponding right hand side of the inequality by one; similarly for “$>$”.

4.1 Weighted Set Multicover

We demonstrate Theorem 3 on the following problem:

[TABLE]

We a ready to prove Theorem 6:

Proof of Theorem 6..

Observe that there are at most $2^{k}$ different sets $F\in 2^{U}$; we classify each pair $(F,w)$ in the input into one of $T\leq 2^{k}$ different types. Moreover, for any two pairs $(F,w)$ and $(F,w^{\prime})$ with $w\leq w^{\prime}$, for any solution containing $(F^{\prime},w^{\prime})$ and not containing $(F,w)$ there is another solution which is at least as good and contains $(F,w)$. We thus order the pairs $(F,w),(F,w^{\prime}),\ldots$ by non-decreasing weight, so that lighter elements will be used before heavier ones in any optimal solution.

This allows us to represent the input instance in a succinct way by $T$ functions $g^{i},\dots,g^{T}:[n]\to\mathbb{N}$ such that, for any $i\in[T]$, $g^{i}(k)$ is defined as the sum of the $k$ lightest elements of type $i$ or $+\infty$ in case that there are less than $k$ elements of type $i$. Observe that since each $g^{i}$ is a partial sum of a non-decreasing sequence of weights, it is a convex function.

We construct a combinatorial $n$-fold IP to solve the problem. Let x_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau} for each {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\in 2^{U} and each $\tau\in[T]$ be a variable. Let l_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau}=u_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau}=0 for each {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\in 2^{U} such that {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}\neq F^{\tau}, and let l_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau}=0 and u_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau}=n for {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}=F^{\tau}. The variable x_{{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}}^{\tau} with {\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}}=F^{\tau} represents the number of sets of type $\tau$ in the solution. The IP formulation then reads

[TABLE]

note that $f_{i}$ is $1$ if i\in{\mathchoice{\mbox{\boldmath\displaystyle\bf f}}{\mbox{\boldmath\textstyle\bf f}}{\mbox{\boldmath\scriptstyle\bf f}}{\mbox{\boldmath\scriptscriptstyle\bf f}}} and [math] otherwise. Let us determine the parameters $\hat{\Delta},\hat{r},\hat{t},\hat{n}$ and $\hat{L}$ of this combinatorial $n$-fold IP instance. Clearly, the largest coefficient $||\hat{D}||_{\infty}$ is $1$, the number of globally uniform constraints $\hat{r}$ is $k$, the number of variables per brick $\hat{t}$ is $2^{k}$, the number of bricks $\hat{n}$ is $T$, and the length of the input $\hat{L}$ is at most $\log n+\log w_{\max}$. ∎

4.2 Stringology

To show that Theorem 3 can be used to speed up many previous result, we show a single-exponential algorithm for an artificial “meta-problem” called $\delta$-Multi Strings which generalizes many previously studied problems:

[TABLE]

We call a distance function $\delta:\Sigma^{*}\times\Sigma^{*}\to\mathbb{N}$ character-wise wildcard-compatible if $\delta(x,y)=\sum_{i=1}^{L}\delta(x[i],y[i])$ for any two strings $x,y\in\Sigma^{L}$, and $\delta(e,\star)=0$ for all $e\in\Sigma$.

Theorem 19**.**

There is an algorithm that solves instances of $\delta$-Multi Strings in time $K^{O(k^{2})}O(\log L)$, where $K=\max\left\{|\Sigma|,k,\max_{e,f\in\Sigma}\delta(e,f)\right\}$ and $\delta$ is any character-wise wildcard-compatible function.

When $\delta$ is the Hamming distance $d_{H}$, it is standard to first “normalize” the input to an equivalent instance over the alphabet $[k]$ [HR15, Lemma 1]. Thus, for $\delta=d_{H}$ we get rid of the dependence on $|\Sigma|$:

Theorem 20**.**

$d_{H}$-Multi Strings can be solved in time $k^{O(k^{2})}O(\log L)$.

Then, Theorem 4 below is a simple corollary of Theorem 20 and the fact that $\delta$-Multi Strings generalizes all the listed problems; see problem definitions and Table 2 below.

Proof of Theorem 19.

Let us fix an instance of $\delta$-Multi Strings. We create an instance of combinatorial $n$-fold IP and show how solving it corresponds to solving the original $\delta$-Multi Strings problem.

As is standard, we represent the input as an $L\times k$ matrix $C$ with entries from $\Sigma\cup\{\star\}$ whose rows are the input strings $s_{1},\dots,s_{k}$. There are at most $T=(|\Sigma|+1)^{k}$ different input column types; let n_{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}} be the number of columns of type {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}\in(\Sigma\cup\{\star\})^{k} and denote ${\cal T}_{c}\subseteq(\Sigma\cup\{\star\})^{k}$ the set of input column types. A solution can be represented as an $L\times(k+1)$ matrix with entries from $\Sigma\cup\{\star\}$ whose last row does not contain any $\star$ symbol. Thus, there are at most $(|\Sigma|+1)^{k}\cdot|\Sigma|$ solution column types {\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}=({\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}},f)\in\big{(}(\Sigma\cup\{\star\})^{k}\times\Sigma\big{)} and we denote {\cal T}_{s}=\big{(}(\Sigma\cup\{\star\})^{k}\times\Sigma\big{)} the set of all solution column types. We say that an input column type {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}\in{\cal T}_{c} is compatible with a solution column type {\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}\in{\cal T}_{s} if {\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}=({\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}},f) for some $f\in\Sigma$.

Let us describe the combinatorial $n$-fold IP formulation. It consists of variables x_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}}^{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}} for each {\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}\in{\cal T}_{s} and each {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}\in{\cal T}_{c}. Intuitively, the variable x_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}}^{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}} encodes the number of columns $\textstyle\bf\alpha$ in the solution; however, to obey the format of combinatorial $n$-fold IP, we need a copy of this variable for each brick, hence the upper index $\textstyle\bf e$. We set an upper bound u_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}}^{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}}=\|{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}\|_{\infty} for each two compatible $\textstyle\bf e$ and $\textstyle\bf\alpha$, and we set u_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}}^{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}}=0 for each pair which is not compatible; all lower bounds are set to [math]. The locally uniform constraints are simply \sum_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}\in{\cal T}_{s}}x_{{\mathchoice{\mbox{\boldmath\displaystyle\bf\alpha}}{\mbox{\boldmath\textstyle\bf\alpha}}{\mbox{\boldmath\scriptstyle\bf\alpha}}{\mbox{\boldmath\scriptscriptstyle\bf\alpha}}}}^{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}}=n_{{\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}} for all {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}\in{\cal T}_{c}. The globally uniform constraints are

[TABLE]

and the objective is

[TABLE]

We then apply Theorem 3 with the following set of parameters:

• •

$\hat{\Delta}$ is one plus the largest coefficient in $D$, which is $\displaystyle{1+\max_{e,f\in\Sigma}\delta(e,f)\leq 1+K}$,

• •

$\hat{r}$ is the number of globally uniform constraints, which is $2k$,

• •

$\hat{t}$ is the number of variables per brick, which is $|{\cal T}_{s}|\leq(|\Sigma|+1)^{k}|\Sigma|\leq K^{2k}$,

• •

$\hat{n}$ is the number of bricks, which is $|{\cal T}_{c}|\leq(|\Sigma|+1)^{k}\leq K^{k}$, and,

• •

$\hat{L}$ is the size of the input \langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},\mathbf{0},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf w}}{\mbox{\boldmath\textstyle\bf w}}{\mbox{\boldmath\scriptstyle\bf w}}{\mbox{\boldmath\scriptscriptstyle\bf w}}}\rangle\leq\log L. ∎

The problem definitions follow. Some problems reduce to solving polynomially (in $L$) or $k^{k}$ many instances of Closest String. In such a case, we mention the fact after introducing the problem, and say that the problem poly-reduces or $\mathsf{FPT}$-reduces to Closest String.

[TABLE]

[TABLE]

[TABLE]

Note. Gramm et al. [GNR03] observe that $d$-Mismatch poly-time reduces to Closest String.

[TABLE]

[TABLE]

Note. Neighbor String is studied by Nishimura and Simjour [NS12]. It generalizes DSS: given an instance of DSS, create an instance of Neighbor String with $d_{i}=0$ and $D_{i}=d_{1}$ for all bad strings, and $d_{i}=d_{2}$ and $D_{i}=L$ for all good strings.

[TABLE]

[TABLE]

Note. Closest to Most Strings is $\mathsf{FPT}$-reducible to Closest String with parameter $k$ [BM11].

[TABLE]

[TABLE]

See Table 2 for a summary of our improvements for the above-mentioned problems.

4.3 Computational social choice

For simplicity, we only show how Theorem 3 can be applied to speed up the $\mathcal{R}$-Swap Bribery for two representative voting rules $\mathcal{R}$. Let us first introduce the necessary definitions and terminology.

Elections. An election $(C,V)$ consists of a set $C$ of candidates and a set $V$ of voters, who indicate their preferences over the candidates in $C$, represented via a preference order $\succ_{v}$ which is a total order over $C$. For ranked candidates $c$ we denote by $\textrm{rank}(c,v)$ their rank in $\succ_{v}$; then $v$’s most preferred candidate has rank 1 and their least preferred candidate has rank $|C|$. For distinct candidates $c,c^{\prime}\in C$, we write $c\succ_{v}c^{\prime}$ if voter $v$ prefers $c$ over $c^{\prime}$. To simplify notation, we sometimes identify the candidate set $C$ with $\{1,\ldots,|C|\}$, in particular when expressing permutations over $C$. We sometimes identify a voter $v$ with their preference order $\succ_{v}$, as long as no confusion arises.

Swaps. Let $(C,V)$ be an election and let $\succ_{v}\in V$ be a voter. For candidates $c,c^{\prime}\in C$, a swap $s=(c,c^{\prime})_{v}$ means to exchange the positions of $c$ and $c^{\prime}$ in $\succ_{v}$; denote the perturbed order by $\succ_{v}^{s}$. A swap $(c,c^{\prime})_{v}$ is admissible in $\succ_{v}$ if $\operatorname{{\rm rank}}(c,v)=\operatorname{{\rm rank}}(c^{\prime},v)-1$. A set $S$ of swaps is admissible in $\succ_{v}$ if they can be applied sequentially in $\succ_{v}$, one after the other, in some order, such that each one of them is admissible. Note that the perturbed vote, denoted by $\succ_{v}^{S}$, is independent from the order in which the swaps of $S$ are applied. We also extend this notation for applying swaps in several votes and denote it $V^{S}$. We specify $v$’s cost of swaps by a function $\sigma^{v}:C\times C\rightarrow\mathbb{Z}$.

Voting rules. A voting rule $\mathcal{R}$ is a function that maps an election $(C,V)$ to a subset $W\subseteq C$ of winners. Let us define two significant classes of voting rules:

Scoring protocols. A scoring protocol is defined through a vector {\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}=(s_{1},\ldots,s_{|C|}) of integers with $s_{1}\geq\cdots\geq s_{|C|}\geq 0$. For each position ${p\in\{1,\ldots,|C|\}}$, value $s_{p}$ specifies the number of points that each candidate $c$ receives from each voter that ranks $c$ as $j^{\text{th}}$ best. Any candidate with the maximum number of points is a winner. Examples of scoring protocols include the Plurality rule with {\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}=(1,0,\ldots,0), the $d$-Approval rule with {\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}=(1,\ldots,1,0,\ldots,0) with $d$ ones, and the Borda rule with {\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}=(|C|-1,|C|-2,\ldots,1,0). Throughout, we consider only natural scoring protocols for which $s_{1}\leq|C|$; this is the case for the aforementioned popular rules.

C1 rules. A candidate $c\in C$ is a Condorcet winner if any other $c^{\prime}\in C\setminus\{c\}$ satisfies $|\{\succ_{v}\in V\mid c\succ_{v}c^{\prime}\}|>|\{v\in V\mid c^{\prime}\succ_{v}c\}|$. A voting rule is Condorcet-consistent if it selects the Condorcet winner in case there is one. Fishburn [Fis77] classified all Condorcet-consistent rules as C1, C2 or C3, depending on the kind of information needed to determine the winner. For candidates $c,c^{\prime}\in C$ let $v(c,c^{\prime})$ be the number of voters who prefer $c$ over $c^{\prime}$, that is, $v(c,c^{\prime})=|\{\succ_{v}\in V\mid c\succ_{v}c^{\prime}\}|$; we write $c<_{M}c^{\prime}$ if $c$ beats $c^{\prime}$ in a head-to-head contest, that is, if $v(c,c^{\prime})>v(c^{\prime},c)$.

A rule $\mathcal{R}$ is C1 if knowing $<_{M}$ suffices to determine the winner, that is, for each pair of candidates $c,c^{\prime}$ we know whether $v(c,c^{\prime})>v(c^{\prime},c),v(c,c^{\prime})<v(c^{\prime},c)$ or $v(c,c^{\prime})=v(c^{\prime},c)$. An example is the Copelandα rule for $\alpha\in[0,1]$, which specifies that for each head-to-head contest between two distinct candidates, if some candidate is preferred by a majority of voters then they obtain one point and the other candidate obtains zero points, and if a tie occurs then both candidates obtain $\alpha$ points; the candidate with largest sum of points wins.

[TABLE]

We say that two voters $v,v^{\prime}$ are of the same type if $\succ_{v}=\succ_{v^{\prime}}$ and $\sigma^{v}=\sigma^{v^{\prime}}$. We are ready to prove Theorem 5:

Proof of Theorem 5..

Let $n_{1},\dots,n_{T}$ be the numbers of voters of given types. Let $x_{j}^{i}$ for $j\in[|C|!]$ and $i\in[T]$ be a variable encoding the number of voters of type $i$ that are bribed to be of order $j$ in the solution. With slight abuse of notation, we denote $\sigma^{i}(i,j)$ the cost of bribery for a voter of type $i$ to change order to $j$ (as by [EFS09, Proposition 3.2] this cost is fixed). Regardless of the voting rule ${\cal R}$, the objective and the locally uniform constraints are identical:

[TABLE]

The number of variables per brick $\hat{t}$ is $|C|!$, the number of bricks $\hat{n}$ is $T$, and the size of the instance $\hat{L}$ is $\log n+\log(|C|^{2}\sigma_{\max})$, because at most $|C|^{2}$ swaps suffice to permute any order $i\in\left[|C|!\right]$ to any other order $j\in[|C|!]$ [EFS09, Proposition 3.2]. Let us now describe the globally uniform constraints separately for the two classes of voting rules which we study here.

Natural scoring protocol.

Let {\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}=(s_{1},\dots,s_{|C|}) be a natural scoring protocol, i.e., $s_{1}\geq\dots\geq s_{|C|}$ and \|{\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}\|_{\infty}\leq|C|. With slight abuse of notation, we denote $s_{j}(c)$, for $j\in[|C|!]$ and $c\in C$, the number of points obtained by candidate $c$ from a voter of order $j$. The globally uniform constraints then enforce that $c^{\star}$ gets at least as many points as any other candidate $c$:

[TABLE]

The number $\hat{r}$ of these constraints is $|C|-1$, and the largest coefficient in them is \|{\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}\|_{\infty}\leq|C| (as is $\textstyle\bf s$ being a natural scoring protocol); therefore, \hat{\Delta}=1+\|{\mathchoice{\mbox{\boldmath\displaystyle\bf s}}{\mbox{\boldmath\textstyle\bf s}}{\mbox{\boldmath\scriptstyle\bf s}}{\mbox{\boldmath\scriptscriptstyle\bf s}}}\|\leq 1+|C|.

Any C1 rule.

Let $\alpha_{j}(c,c^{\prime})$ be $1$ if a voter with order $j\in[|C|!]$ prefers $c$ to $c^{\prime}$ and [math] otherwise. Recall that a voting rule is C1 if, to determine the winner, it is sufficient to know, for each pair of candidates $c,c^{\prime}$, whether $v(c,c^{\prime})>v(c^{\prime},c),v(c,c^{\prime})<v(c^{\prime},c)$ or $v(c,c^{\prime})=v(c^{\prime},c)$; where $v(c,c^{\prime})=|\{v\mid c\succ_{v}c^{\prime}\}|$. We call a tuple $<_{M}\in\{<,=,>\}^{|C|^{2}}$ a scenario. Thus, a C1 rule can be viewed as partitioning the set of all scenarios into those that select $c^{\star}$ as a winner and those that do not. Then, it suffices to enumerate all the at most $3^{|C|^{2}}$ scenarios $<_{M}$ where $c^{\star}$ wins, and for each of them to solve a combinatorial $n$-fold IP with globally uniform constraints enforcing the scenario $<_{M}$:

[TABLE]

The number $\hat{r}$ of these constraints is $\binom{|C|}{2}\leq|C|^{2}$, and the largest coefficient in them is 1, so $\hat{\Delta}=2$. The proof is finished by plugging in the values $\hat{\Delta},\hat{r},\hat{t},\hat{n}$ and $\hat{L}$ into Theorem 3. ∎

4.4 Huge $n$-fold integer programming with small domains

Onn introduces the Huge $n$-fold integer programming problem [Onn14, OS15], which concerns problems that can be formulated as an $n$-fold IP with the number of bricks $n$ given in binary. Bricks are thus represented not explicitly, but succinctly by their multiplicity. It is at first unclear if this problem admits an optimal solution which can be encoded in polynomial space, but this is possible by a theorem of Eisenbrand and Shmonin [ES06, Theorem 2], as pointed out by Onn [Onn14, Theorem 1.3 (1)]. Thus, the problem is as follows:

Let $E=\left(\begin{smallmatrix}D\\ A\end{smallmatrix}\right)\in\mathbb{Z}^{(r+s)\times t}$ be a bimatrix. Let $T$ be a positive integer representing the number of types of bricks. We are given $T$ positive integers $n_{1},\dots,n_{T}$ with $n=\sum_{i=1}^{T}n_{i}$ and vectors {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{0}\in\mathbb{Z}^{r} and {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}^{i},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}^{i}\in\mathbb{Z}^{t} and {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{i}\in\mathbb{Z}^{s}, and a separable convex function $f^{i}$, for every $i\in[T]$. For $i\in[T]$ and $\ell\in[n_{i}]$ we define the index function $\iota$ as $\iota(i,\ell):=\left(\sum_{j=1}^{i-1}n_{j}\right)+\ell$.

We call an $n$-fold IP instance given by the constraint matrix $E^{(n)}$ with the right hand side \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}} defined by \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}}^{\iota(i,\ell)}:={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{i} and \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}}^{0}:={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{0}, lower and upper bounds defined by \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}}^{\iota(i,\ell)}:={\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}^{i} and \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}^{\iota(i,\ell)}:={\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}^{i} with the objective function f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}):=\sum_{i=1}^{n}\sum_{\ell=1}^{m_{i}}f^{i}\big{(}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{\iota(i,\ell)}\big{)} the huge instance.

For $i\in[T]$ and $\ell\in[n_{i}]$, and a feasible solution $\textstyle\bf x$ of the huge instance, we say that the brick {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{\iota(i,\ell)} is of type $i$, and we say that {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{\iota(i,\ell)} has configuration {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in\mathbb{Z}^{t} with \hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}}^{i}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\leq\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}}^{i} if {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{\iota(i,\ell)}={\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}. The succinct representation of $\textstyle\bf x$ is the set of tuples \left\{\left({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}^{i,j},m^{i,j}\right)\mid\mbox{{\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}$has$m^{i,j}$bricks of type$i$with configuration${\mathchoice{\mbox{\boldmath$\displaystyle\bf x$}}{\mbox{\boldmath$\textstyle\bf x$}}{\mbox{\boldmath$\scriptstyle\bf x$}}{\mbox{\boldmath$\scriptscriptstyle\bf x$}}}^{i,j}}\right\}.

[TABLE]

In the special case of small domains we obtain the following:

Theorem 21**.**

Let $d_{1},\dots,d_{t}\in\mathbb{N}$ be such that $d_{j}=\max_{i\in[n]}u_{j}^{i}-l_{j}^{i}$, $d_{\max}=\max_{j\in[t]}d_{j}$ and let $\delta=\prod_{j=1}^{t}d_{j}$. Then the huge $n$-fold IP problem can be solved in time $\delta^{O(r)}(td_{\max}\|D\|_{\infty}r)^{O(r^{2})}O(T^{3}\log n)$.222In fact, our result holds even in the case when $f$ is an arbitrary (i.e. non-convex) function, but this does not imply any more power because of bounded domains.

This result is useful for the following reason. Knop et al. [KKM17b] obtain $n$-fold IP formulations with small domains for the very general ${\cal R}$-Multi Bribery problem. Together with Theorem 21, this immediatelly implies an exponential speedup in the number of bricks, without having to reformulate these problems as combinatorial $n$-fold IPs. However, there are still benefits in using Theorem 3 directly (as shown in previous sections) as it leads to better dependence on the respective parameters.

Proof of Theorem 21.

Let E,n_{1},\dots,n_{T},{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f be an instance $I$ of Huge $n$-fold IP. First, we shall prove that we can restrict our attention to the case where {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}=\mathbf{0} and {\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}^{i}\leq(d_{1},\dots,d_{t})={\mathchoice{\mbox{\boldmath\displaystyle\bf d}}{\mbox{\boldmath\textstyle\bf d}}{\mbox{\boldmath\scriptstyle\bf d}}{\mbox{\boldmath\scriptscriptstyle\bf d}}} for all $i\in[T]$. Consider a variable $x_{j}^{i}$ with $l_{j}^{i}\neq 0$ and any row {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}^{\intercal}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}=b of of the system E^{(n)}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}. Because the contribution of $x_{j}^{i}$ to the right hand side is $e_{j}^{i}x_{j}^{i}$, we have that

[TABLE]

Let $I^{\prime}$ be an instance of Huge $n$-fold IP obtained from $I$ by, for every row {\mathchoice{\mbox{\boldmath\displaystyle\bf e}}{\mbox{\boldmath\textstyle\bf e}}{\mbox{\boldmath\scriptstyle\bf e}}{\mbox{\boldmath\scriptscriptstyle\bf e}}}{\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}=b and every variable $x_{j}^{i}$, changing the right hand side from $b$ to $b-e_{j}^{i}l_{j}^{i}$, and setting $u_{j}^{i}:=u_{j}^{i}-l_{j}^{i}$ and $l_{j}^{i}:=0$. Clearly there is a bijection between the feasible solutions of $I$ and $I^{\prime}$ such that if {\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}} is a feasible solution of $I^{\prime}$, $\textstyle\bf x$ is a feasible solution of $I$, and thus minimizing f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}-{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}) over $I^{\prime}$ is equivalent to minimizing f({\mathchoice{\mbox{\boldmath\displaystyle\bf x}}{\mbox{\boldmath\textstyle\bf x}}{\mbox{\boldmath\scriptstyle\bf x}}{\mbox{\boldmath\scriptscriptstyle\bf x}}}) over $I$. Thus, from now on assume that $I$ satisfies {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}}=\mathbf{0} and {\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}^{i}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf d}}{\mbox{\boldmath\textstyle\bf d}}{\mbox{\boldmath\scriptstyle\bf d}}{\mbox{\boldmath\scriptscriptstyle\bf d}}} for all $i\in[T]$.

Let ${\cal C}^{i}$ for $i\in[T]$ be the set of all possible configurations of a brick of type $i$, defined as {\cal C}^{i}=\left\{{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in\mathbb{Z}^{t}\mid A{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}={\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{i},\mathbf{0}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\leq{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}^{i}\right\} and let ${\cal C}=\prod_{j=1}^{t}[0\,:\,d_{j}]$ be the set of all configurations. Clearly, ${\cal C}^{i}\subseteq{\cal C}$ for all $i\in[T]$, and $|{\cal C}|=\delta$. Let $C\in\mathbb{Z}^{t\times\delta}$ be a matrix whose columns are all configurations from ${\cal C}$.

We shall give a combinatorial $n$-fold IP formulation solving the huge $n$-fold IP instance $I$. The formulation contains variables y_{{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}}^{i} for each {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in{\cal C} and each $i\in[T]$ encoding how many bricks of type $i$ have configuration $\textstyle\bf c$ in the solution of $I$. The formulation then is

[TABLE]

It remains to verify that the formulation above corresponds to the huge $n$-fold IP instance $I$. The objective (11) clearly has the same value. Consider the globally uniform constraints (12). In the huge $n$-fold IP instance, a configuration {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in{\cal C}^{i} of a brick of type $i\in[T]$ contributes D{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}} to the right hand side in the first $r$ rows. This corresponds in our program to the column D{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}} of the matrix $DC$. The locally uniform constraints (13) simply state that the solution needs to contain exactly $n_{i}$ bricks of type $i$. Finally, since a brick of type $i$ can never have a configuration {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\not\in{\cal C}^{i} we set all variables y_{{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}}^{i} with {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\not\in{\cal C}^{i} to zero with the upper bound (14), and place no restrictions on y_{{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}}^{i} with {\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in{\cal C}^{i} (15).

The parameters $\hat{\Delta},\hat{r},\hat{t},\hat{n},\hat{L}$ of the resulting combinatorial $n$-fold IP are:

• •

$\hat{\Delta}$ is one plus the largest coefficient in the upper matrix $\hat{D}=DC$, which is $\|DC\|_{\infty}\leq td_{\max}\|D\|_{\infty}$,

• •

the number of globally uniform constraints $\hat{r}=r$,

• •

the number of variables in a brick $\hat{t}=\delta$,

• •

the number of bricks $\hat{n}=T$, and,

• •

the input length \hat{L}=\langle\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}},\mathbf{0},\hat{{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}},\hat{f}\rangle\leq\log n\cdot\left(\max_{i\in[T]}\max_{{\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}}\in{\cal C}^{i}}f^{i}({\mathchoice{\mbox{\boldmath\displaystyle\bf c}}{\mbox{\boldmath\textstyle\bf c}}{\mbox{\boldmath\scriptstyle\bf c}}{\mbox{\boldmath\scriptscriptstyle\bf c}}})\right). ∎

4.5 Miscellaneous

Recall from Sect. 1 our comparison of the Gramm et al. [GNR03] ILP for Closest String to a similar combinatorial $n$-fold IP:

[TABLE]

where $D=(D_{1}\leavevmode\nobreak\ D_{2}\leavevmode\nobreak\ \dots\leavevmode\nobreak\ D_{k^{k}})$.

This similarity strongly suggests a general way how to construct the formulation on the right given the formulation on the left. Since formulations like the one on the left are ubiquitous in the literature, this would immediatelly imply exponential speed-ups for all such problems.

Definition 22** (Combinatorial pre-$n$-fold IP).**

Let $T,r,t_{1},\dots,t_{T}\in\mathbb{N}$ and $D_{i}\in\mathbb{Z}^{r\times t_{\tau}}$ for each $\tau\in[T]$. Let $t=t_{1}+\cdots t_{T}$ and $D=(D_{1}\cdots D_{T})$. Let

[TABLE]

Moreover, let {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}=({\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{0},b^{1},\dots,b^{T}) with {\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}}^{0}\in\mathbb{Z}^{r} and $b^{\tau}\in\mathbb{Z}$ for each $\tau\in[T]$, {\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}}\in\mathbb{Z}^{t}, and let $f:\mathbb{Z}^{t}\to\mathbb{Z}$ be a separable convex function. Then for $\diamondsuit\in\{<,\leq,=,\geq,>\}^{r+T}$, a combinatorial pre-$n$-fold IP is the problem

[TABLE]

Corollary 23**.**

Any combinatorial pre-$n$-fold IP with L=\langle{\mathchoice{\mbox{\boldmath\displaystyle\bf b}}{\mbox{\boldmath\textstyle\bf b}}{\mbox{\boldmath\scriptstyle\bf b}}{\mbox{\boldmath\scriptscriptstyle\bf b}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf l}}{\mbox{\boldmath\textstyle\bf l}}{\mbox{\boldmath\scriptstyle\bf l}}{\mbox{\boldmath\scriptscriptstyle\bf l}}},{\mathchoice{\mbox{\boldmath\displaystyle\bf u}}{\mbox{\boldmath\textstyle\bf u}}{\mbox{\boldmath\scriptstyle\bf u}}{\mbox{\boldmath\scriptscriptstyle\bf u}}},f\rangle and $\Delta=1+\|D\|_{\infty}$ can be solved in time $t^{O(r)}(\Delta r)^{O(r^{2})}O(n^{3}L)+\mathsf{oo}$, where $\mathsf{oo}$ is the time required by one call to an optimization oracle of the continuous relaxation of the given IP.

Proof.

We shall create a combinatorial $n$-fold IP instance based on the input combinatorial pre-$n$-fold IP. Let

[TABLE]

Let $\bar{T}_{i}=\sum_{j=1}^{\tau}t_{\tau}$ for each $\tau\in[T]$. Let \varphi:\bigcup_{\tau=1}^{T}\big{(}\{\tau\}\times[t_{\tau}]\big{)}\to[T]\times[t] be an injective mapping from the original variables to the new ones, defined as follows: for each $\tau\in[T]$ and $j\in[t_{\tau}]$, $\varphi(\tau,j)=(\tau,\bar{T}_{\tau-1}+j)$. We call any pair $(\tau,j)\in\mathbb{N}^{2}$ without a preimage in $\varphi$ dummy.

Then, for any $\tau\in[T]$ and $j\in[t_{\tau}]$, let $(\tau^{\prime},j^{\prime})=\varphi(\tau,j)$, and set $\hat{l}^{\tau^{\prime}}_{j^{\prime}}=l^{\tau}_{j}$, $\hat{u}^{\tau^{\prime}}_{j^{\prime}}=u^{\tau}_{j}$, and $\hat{f}^{\tau^{\prime}}_{j^{\prime}}=f^{\tau}_{j}$. For any $\tau\in[T]$ and $j\in[t]$ which form a dummy pair, set $\hat{u}^{\tau}_{j}=\hat{l}^{\tau}_{j}=0$ and let $\hat{f}^{\tau}_{j}$ be the zero function.

Now we see that

[TABLE]

is a combinatorial $n$-fold IP (with inequalities) and thus can be solved in the claimed time by Theorem 3. ∎

Parameterizing by the number of numbers.

Fellows et al. [FGR12] argue that many problems’ $\mathsf{NP}$-hardness construction is based on having many distinct objects, when it would be quite natural that the number of distinct objects, and thus the numbers representing them, is bounded by a parameter. Fellows et al. [FGR12, Theorem 5] give a fixed-parameter algorithm for a certain Mealy automaton problem which models, for example the Heat-sensitive scheduling problem parameterized by the number of distinct heat levels. Their algorithm relies on Lenstra’s algorithm and has a doubly-exponential dependence on the parameter. In the conclusions of their paper, the authors state that “[o]ur main FPT result, Theorem 5, has a poor worst-case running-time guarantee. Can this be improved—at least in important special cases?”

It is easy to observe that their ILP only has $\sigma$ constraints, and that the coefficients are bounded by $|S|^{|S|^{2}}$. Thus, applying our algorithm with $t=\Delta=O(|S|^{|S|^{2}})$ and $r=\sigma$ yields an algorithm with runtime $|S|^{O(\sigma|S|^{2})}$, which is single-exponential in their (combined) parameter.

Scheduling meets fixed-parameter tractability.

Mnich and Wiese [MW15] study the parameterized complexity of fundamental scheduling problems, starting with Makespan Minimization on identical machines. They give an algorithm with doubly-exponential dependence for parameter maximum job length $p_{\max}$, and polylogarithmic dependence on the number $m$ of machines. Knop and Koutecký [KK17] use standard $n$-fold IP to reduce the dependence on $p_{\max}$ to single-exponential, however their algorithm depends polynomially on $m$.

We observe that in the proof of [MW15, Theorem 2] there is an ILP related to a set of configurations which is of size $p_{\max}^{p_{\max}}$. The coefficients of this ILP are unbounded, but by [MW15, Lemma 1] we know that all the coefficients differ by at most $p_{\max}^{p_{\max}}$. Moreover, because the number of machines $m$ is fixed before the construction of the ILP, we can appropriately subtract from the right hand sides and decrease all coefficients such that $\Delta\leq p_{\max}^{p_{\max}}$. Then, take the constraints (8) as globally uniform and notice that there are $r=p_{\max}$ of them. In conclusion, we have $t=\Delta\leq p_{\max}^{p_{\max}}$ and $r=p_{\max}$, which yields an algorithm with run time $p_{\max}^{p_{\max}^{2}}\cdot(\log n+\log m)$; we therefore improve both the algorithm by Mnich and Wiese [MW15] as well as the algorithm by Knop and Koutecký [KK17].

Lobbying in multiple referenda.

Bredereck et al. [BCH*+*14] study the computational social choice problem of lobbying in multiple referenda, and show a Lenstra-based fixed-parameter algorithm for the parameter $m$=“number of choices”. The number of choices induces the parameter $\ell$=“number of ballots”, which is clearly bounded by $\ell\leq 2^{m}$. This leads to an ILP formulation with $2^{O(m)}$ variables, and thus to a run time double-exponential in $m$.

We point out that the proof of [BCH*+*14, Theorem 9] contains a combinatorial pre-$n$-fold IP, with one set of constraints indexed over $i$, and another set of constraints indexed by $j$, which are simply sums of variables over all $i$. Since $i\in[m]$ and $j\in[\ell]$, we have the parameters $\Delta=1$, $t=2^{m}$, $r=m$. Therefore, we improve their algorithm to only single-exponential in $m$, namely $m^{O(m^{2})}\log n$.

Weighted set multicover (WSM) in graph algorithms.

In Subsect. 4.1, we give a combinatorial $n$-fold IP formulation for WSM. While WSM was studied in the context of computational social choice by Bredereck et al. [BFN*+*15], it appeared implicitly several times in algorithms for restricted classes of graphs, namely graphs of bounded vertex cover number and neighborhood diversity. We briefly mention some of these results which we improve here.

Fiala et al. [FGK11] show that Equitable Coloring and $L(p,1)$-Coloring parameterized by vertex cover are fixed-parameter tractable. In the proof of [FGK11, Theorem 4] and similarly of [FGK11, Theorem 9] they construct a combinatorial pre-$n$-fold IP. Lampis [Lam12] introduced the neighborhood diversity parameter and used combinatorial pre-$n$-folds to show that Graph Coloring and Hamiltonian Cycle are $\mathsf{FPT}$ with this parameterization. Ganian [Gan12] later showed in a similar fashion that Vertex Disjoint Paths and Precoloring Extension are also fixed-parameter tractable parameterized by neighborhood diversity. Similarly, Fiala et al. [FGK*+*17] shows that the Uniform Channel Assignment problem is (triple-exponential) fixed-parameter tractable parameterized by neighborhood diversity and the largest edge weight. Ganian and Obdržálek [GO16] and later Knop et al. [KKMT17] study extensions of the MSO logic, and provide fixed-parameter algorithms for their model checking on graphs of bounded neighborhood diversity. These algorithms again use combinatorial pre-$n$-folds under the hood.

We can speed up all aforementioned algorithms by applying Corollary 23.

5 Discussion

We established new and fast fixed-parameter algorithms for a class of $n$-fold IPs, which led to the first single-exponential time algorithms for several well-studied problems. Many intriguing questions arise; e.g., is Huge $n$-fold IP fixed-parameter tractable for parameter $(r,s,t,\Delta)$? One sees that optimality certification is fixed-parameter tractable using ideas similar to Onn [Onn14]; yet, one possibly needs exponentially (in the input size) many augmenting steps.

For most of our applications, complexity lower bounds are not known to us. Our algorithms yield complexity upper bounds of $k^{O(k^{2})}$ on the dependence on parameter $k$ for various problems, such as Closest String, Weighted Set Multicover, Score-Swap Bribery or even Makespan Minimization [KK17]. Is this just a common feature of our algorithm, or are there hidden connections between some of these problems? And what are their actual complexities? All we know so far is a trivial ETH-based $2^{o(k)}$ lower bound for Closest String based on its reduction from Satisfiability [FL97].