The Complexity Landscape of Decompositional Parameters for ILPJournal: Artificial Intelligence

Let $\alpha$ be a solution of $\text{var}(I)$ of value $w=\eta(\alpha)$. Since ${\cal F}^{\prime}\subseteq{\cal F}$, it follows that setting $\alpha^{\prime}$ to be a restriction of $\alpha$ to $\text{var}(I)\setminus T_{y}$ satisfies every inequality in ${\cal F}^{\prime}$. Since variables in $T_{y}$ do not contribute to $\eta$, it also follows that $\eta(\alpha)=\eta(\alpha^{\prime})$.

On the other hand, let $\alpha^{\prime}$ be a solution of $\text{var}(I^{\prime})$ of value $w=\eta^{\prime}(\alpha^{\prime})$. Consider the assignment $\alpha$ obtained by extending $\alpha^{\prime}$ to $T_{y}$ by reusing the assignments of $T_{x}$ on $T_{y}$. Formally, for each $z\in T_{y}$ we set $\alpha(z)=\alpha^{\prime}(\delta^{-1}_{x,y}(z))$ and for all other variables $w\in\text{var}(I^{\prime})$ we set $\alpha(w)=\alpha^{\prime}(w)$. By assumption, $\alpha$ and $\alpha^{\prime}$ must assign the same values to any variable $w$ such that $s_{w}\neq 0$, and hence $\eta(\alpha)=\eta(\alpha^{\prime})$. To argue feasibility, first observe that any $A\in{\cal F}^{\prime}$ must be satisfied by $\alpha$ since $\alpha$ and $\alpha^{\prime}$ only differ on variables which do not occur in $I^{\prime}$. Moreover, by definition of $\sim$ for each $A\in{\cal F}\setminus{\cal F}^{\prime}={\cal F}(T_{y})$ there exists an inequality $A^{\prime}\in{\cal F}^{\prime}$ such that $\delta_{x,y}(A^{\prime})=A$. In particular, this implies that $A(\alpha)=A^{\prime}(\alpha)=A^{\prime}(\alpha^{\prime})$, and since $A^{\prime}(\alpha^{\prime})\leq b_{A^{\prime}}=b_{A}$ we conclude that $A(\alpha)\leq b_{A}$. Consequently, $\alpha$ satisfies $A$.

The final claim of the lemma follows from the construction of $\alpha$ described above. ∎

Consider an element $a\in Z$. By construction of $G_{I}$, each inequality $A\in{\cal F}(T_{a})$ only contains at most $k-i$ variables outside of $T_{a}$ (specifically, the ancestors of $a$) and at most $i$ variables in $T_{a}$. Furthermore, $b_{A}$ and each coefficient of a variable in $A$ is an integer whose absolute value does not exceed $\ell$. From this it follows that there exists a finite number of inequalities which can occur in ${\cal F}(T_{a})$. Specifically, the number of distinct combinations of coefficients for all the variables in $A$ and for $b_{A}$ is $(2\ell+1)^{k+1}$, and the number of distinct choices of variables in $\text{var}(A)\cap T_{a}$ is upper-bounded by ${m\choose i}$, and so we arrive at $|{\cal F}(T_{a})|\leq(2\ell+1)^{k+1}\cdot{m\choose i}\leq(2\ell+1)^{k+1}\cdot m^{i}$.

Consequently, the set of inequalities for each child $y\in Z$ of $z$ has bounded cardinality. We will use this to bound the number of equivalence classes in $\textup{\#C}(\ell,k,i,m)$ by observing that two elements are equivalent if and only if they occur in precisely the same sets of inequalities (up to renaming). To formalize this intuition, we need a formal way of canonically renaming all variables in the individual subtrees rooted in $Z$; without renaming, each ${\cal F}(T_{y})$ would span a distinct set of variables and hence it would not be possible to bound the set of all such inequalities. So, for each $y$ let $\delta_{y,x_{0}}$ be a bijective renaming function which renames all of the variables in $T_{y}$ to the variable set $\{x_{0}^{1},x_{0}^{2},\dots,x_{0}^{|T_{y}|}\}$ (in an arbitrary way). Now we can formally define $\Gamma_{z}=\{\,{\cal F}(T_{x_{0}})\;{:}\;\delta_{y,x_{0}}({\cal F}(T_{y})),y\in Z\,\}$, and observe that $\Gamma_{z}$ has cardinality at most $2^{(2\ell+1)^{k+1}\cdot m^{i}}=\textup{\#C}(\ell,k,i,m)$. To conclude the proof, recall that if two variables $a,b$ satisfy ${\cal F}(T_{a})=\delta_{b,a}({\cal F}(T_{b}))$ for a bijective renaming function $\delta_{b,a}$, then $b\sim a$. Hence, the absolute bound on the cardinality of $\Gamma_{z}$ implies that $\sim$ has at most $\textup{\#C}(\ell,k,i,m)$ equivalence classes over $Z$. ∎

Consider the following algorithm $\mathbb{A}$. First, $\mathbb{A}$ computes a subset $Z^{\prime}$ consisting of exactly (arbitrarily chosen) $\textup{\#C}(\ell,k,i,m)+1$ children of $Z$. Then $\mathbb{A}$ branches over all distinct pairs $x,y\in Z^{\prime}$ in time at most $\mathcal{O}(\textup{\#C}(\ell,k,i,m)^{2})$. Second, $\mathbb{A}$ branches over all of the at most $m!$ bijective renaming functions $\delta_{x,y}$. Third, $\mathbb{A}$ computes $\delta_{x,y}({\cal F}(T_{x}))$ and tests whether it is equal to ${\cal F}(T_{y})$ (which takes at most $\mathcal{O}(m)$ time); if this is the case, then $\mathbb{A}$ terminates and outputs $x,y$ and $\delta_{x,y}$.

We argue correctness. By Lemma 8 and due to the cardinality of $Z^{\prime}$, there must exist $x,y\in Z^{\prime}$ such that $x\sim y$. In particular, there must exist a renaming function $\delta_{x,y}$ such that $\delta_{x,y}({\cal F}(T_{x}))={\cal F}(T_{y})$. But then $\mathbb{A}$ is guaranteed to find such $x,y,\delta_{x,y}$ since it performs an exhaustive search. ∎

In order to avoid having to consider all children of $z$, the algorithm first computes (an arbitrary) subset $Z^{\prime}$ of $Z$ such that $|Z^{\prime}|=\textup{\#C}(\ell,k,i,m)+2$. Then to be able to apply Lemma 9 without changing the set of solutions of $I$, the algorithm computes a subset $Z^{\prime\prime}$ of $Z^{\prime}$ such that $|Z^{\prime\prime}|=\textup{\#C}(\ell,k,i,m)+1$ and for every $z^{\prime}\in Z^{\prime\prime}$ it holds that $s_{z^{\prime\prime}}=0$ for every $z^{\prime\prime}\in T_{z^{\prime}}$. Note that since there are at most $k$ variables of $I$ with non-zero coefficients in $\eta$ and these variables form a clique in $G_{I}$, all of them occur only in a single branch of $T_{z}$. It follows that $Z^{\prime\prime}$ as specified above exists and it can be obtained from $Z^{\prime}$ by removing the (at most one) element $z^{\prime}$ in $Z^{\prime}$ with $s_{z^{\prime\prime}}\neq 0$ for some $z^{\prime\prime}\in T_{z^{\prime}}$. Observe that this step of the algorithm takes time at most $\mathcal{O}(m\cdot(\textup{\#C}(\ell,k,i,m)+1))$.

The algorithm then proceeds as follows. It uses Lemma 9 to find two variables $x,y\in Z^{\prime\prime}$ such that $x\sim y$ and computes $I^{\prime}$ from $I$ by omitting $T_{y}$ from $I$. The running time of the algorithm follows from Lemma 9 since the running times of the other steps of the algorithm are dominated by the application of Lemma 9. The corollary now follows from Lemma 7 and Lemma 9, which certify that:

• •

  there exists a solution $\alpha$ of $I$ of value $w=\eta(\alpha)$ if and only if there exists a solution $\alpha^{\prime}$ of $I^{\prime}$ of value $w$, and

• •

  a solution $\alpha$ can be computed from any solution $\alpha^{\prime}$ in linear time. ∎

The algorithm exhaustively applies Corollary 10 to every variable of $T$ in a bottom-up manner, i.e., it starts by applying the corollary exhaustively to all variables at depth $k-1$ and then proceeds up the levels of $T$ until it reaches depth $1$. Let $T^{\prime}$ be the subtree of $T$ obtained after the exhaustive application of Corollary 10 to $T$.

We will first show that if $x$ is a variable at depth $i$ of $T^{\prime}$, then $x$ has at most $d_{i}$ children and $|T_{x}^{\prime}|\leq e_{i}$. We will show the claim by induction on the depth $i$ starting from depth $k$. Because all variables $x$ of $T$ at level $k$ are leaves, it holds that $x$ has $0=d_{k}$ children in $T^{\prime}$ and $|T_{x}^{\prime}|=1\leq e_{k}$, showing the start of the induction. Now let $x$ be a variable at depth $i$ of $T^{\prime}$ and let $y$ be a child of $x$ in $T^{\prime}$. It follows from the induction hypothesis that $|T_{y}^{\prime}|\leq e_{i+1}$. Moreover, using Corollary 10, we obtain that $x$ has at most $\tilde{\textup{\#C}}(\ell,k,i,e_{i+1})+1=d_{i}$ children in $T^{\prime}$ and thus $|T_{x}^{\prime}|\leq d_{i}e_{i+1}+1=e_{i}$, as required.

The running time of the algorithm now follows from the observation that (because every application of Corollary 10 removes at least one variable of $I$) Corollary 10 is applied at most $|I|$ times and moreover the maximum running time of any call to Corollary 10 is at most $\mathcal{O}(d_{1}^{2}\cdot e_{1}!e_{1})$. Correctness and the fact that $\alpha$ can be computed from $\alpha^{\prime}$ follow from Corollary 10; more specifically, we extend $\alpha^{\prime}$ into $\alpha$ by assigning pruned variables in the same way as their equivalent counterparts. ∎

The algorithm proceeds in three steps. First, it applies Lemma 11 to reduce the instance $I$ into an “equivalent” instance $I^{\prime}$ containing at most $e_{1}$ variables in time ${\mathcal{O}}(|I|d_{1}^{2}\cdot e_{1}!e_{1})$; in particular, a solution $\alpha$ of $I$ can be computed in linear time from a solution $\alpha^{\prime}$ of $I^{\prime}$. Second, it uses Theorem 1 to compute a solution $\alpha^{\prime}$ of $I^{\prime}$ in time at most $\mathcal{O}(e_{1}^{2.5e_{1}+o(e_{1})}\cdot|I^{\prime}|)$; because $e_{1}$ and $d_{1}$ are bounded by our parameters, the whole algorithm runs in FPT time. Third, it transforms the solution $\alpha^{\prime}$ into a solution $\alpha$ of $I$. Correctness follows from Lemma 11 and Theorem 1. ∎

We will show the theorem by a polynomial-time reduction from the well-known NP-hard $3$-Colorability problem [16]: given a graph, decide whether the vertices of $G$ can be colored with three colors such that no two adjacent vertices of $G$ share the same color.

The main idea behind the reduction is to represent a $3$-partition of the vertex set of $G$ (which in turn represents a $3$-coloring of $G$) by the domain values of three “global” variables. The value of each of these global variables will represent a subset of vertices of $G$ that will be colored using the same color. To represent a subset of the vertices of $G$ in terms of domain values of the global variables, we will represent every vertex of $G$ with a unique prime number and a subset by the value obtained from the multiplication of all prime numbers of vertices contained in the subset. To ensure that the subsets represented by the global variables correspond to a valid $3$-partition of $G$ we will introduce constraints which ensure that:

• C1

  For every prime number representing some vertex of $G$ exactly one of the global variables is divisible by that prime number. This ensures that every vertex of $G$ is assigned to exactly one color class.

• C2

  For every edge $\{u,v\}$ of $G$ it holds that no global variable is divisible by the prime numbers representing $u$ and $v$ at the same time. This ensures that no two adjacent vertices of $G$ are assigned to the same color class.

Thus let $G$ be the given instance of $3$-Coloring and assume that the vertices of $G$ are uniquely identified as elements of $\{1,\dots,|V(G)|\}$. In the following we denote by $p(i)$ the $i$-th prime number for any positive integer $i$, where $p(1)=2$. We construct an instance $I$ of ILP-feasibility in polynomial-time with treedepth at most $8$ and coefficients bounded by a polynomial in $V(G)$ such that $G$ has a $3$-coloring if and only if $I$ has a feasible assignment. This instance $I$ has the following variables:

• •

  The global variables $g_{1}$, $g_{2}$, and $g_{3}$ with an arbitrary positive domain, whose values will represent a valid $3$-Partioning of $V(G)$.

• •

  For every $i$ and $j$ with $1\leq i\leq|V(G)|$ and $1\leq j\leq 3$, the variables $m_{i,j}$ (with an arbitrary non-negative domain), $r_{i,j}$ (with domain between $0$ and $p(i)-1$), and $u_{i,j}$ (with binary domain). These variables are used to secure condition C1.

• •

  For every $e\in E(G)$, $v\in e$, and $j$ with $1\leq j\leq 3$, the variables $m_{e,v,j}$ (with an arbitrary non-negative domain), $r_{e,v,j}$ (with domain between $0$ and $p(v)-1$), and $u_{e,v,j}$ (with binary domain). These variables are used to secure condition C2.

Following the construction and explanations provided above, it is not difficult to see that $I$ has a feasible assignment if and only if $G$ has a $3$-coloring. Indeed, for any $3$-coloring of $G$, one can construct a feasible assignment of $I$ by computing the prime-number encoding for the vertex sets that receive colors $1,2,3$ and assign these three numbers to $g_{1},g_{2},g_{3}$, respectively. Such an assignment allows us to straightforwardly satisfy the constraints ensuring C1 holds (since each prime occurs in exactly one global constraint), the constraints ensuring C2 holds (since each edge is incident to at most one of each color) while maintaining the domain bounds.

On the other hand, for any feasible assignment $\alpha$, clearly each of $\alpha(g_{1}),\alpha(g_{2}),\alpha(g_{3})$ will be divisible by some subset of prime numbers between $2$ and $p(|V(G)|)$. In particular, since $\alpha$ is feasible it follows from the construction of our first group of constraints that each prime between $2$ and $p(|V(G)|)$ divides precisely one of $\alpha(g_{1}),\alpha(g_{2}),\alpha(g_{3})$, and so this uniquely encodes a corresponding candidate $3$-coloring for the vertices of the graph. Finally, since $\alpha$ also satisfies the second group of constraints, this candidate $3$-coloring must have the property that each edge is incident to exactly $2$ colors, and so it is in fact a valid $3$-coloring.

It remains to show that the treedepth of $I$ is at most $8$. We will show this by using the characterization of treedepth given in Proposition 4. We first observe that the graph $G_{I}\setminus\{g_{1},g_{2},g_{3}\}$ consists of the following components:

• •

  for every $i$ with $1\leq i\leq|V(G)|$, one component on the vertices $m_{i,1},\dotsc,m_{i,3}$, $r_{i,1},r_{i,2},r_{i,3}$, $u_{i,1},u_{i,2},u_{i,3}$. Note that all of these components are isomorphic to each other and we will therefore in the following refer to these components as vertex-type components.

• •

  for every $e=\{w,v\}\in E(G)$ and $j$ with $1\leq j\leq 3$, one component on the vertices $m_{e,w,j}$, $m_{e,v,j}$, $r_{e,w,j}$, $r_{e,v,j}$, $u_{e,w,j}$, and $u_{e,v,j}$. Note that all of these components are isomorphic to each other and we will therefore in the following refer to these components as edge-type components.

The two types of components are illustrated in Figure 1. We will show next that any vertex-type component has treedepth at most $5$ and every edge-type component has treedepth at most $4$. This would then imply that $G_{I}$ has treedepth at most $8$ (since it suffices to remove the vertices $\{g_{1},g_{2},g_{3}\}$ in order to decompose the graph into these components). Hence let $i$ with $1\leq i\leq|V(G)|$ and consider the vertex-type component $C_{i}$ on the vertices $m_{i,1},m_{i,2},m_{i,3}$, $r_{i,1},r_{i,2},r_{i,3}$, $u_{i,1},u_{i,2},u_{i,3}$. Note that $C_{i}\setminus\{u_{i,1},u_{i,2},u_{i,3}\}$ consists of one component for every $j$ with $1\leq j\leq 3$ that contains the vertices $m_{i,j}$ and $r_{i,j}$. Clearly each of these three components has treedepth at most $2$ and hence the treedepth of $C_{i}$ is at most $2+3=5$, as required.