Incremental PCST: Algorithmic Insights

• Incremental PCST is a variant where edges are acquired in sequence ensuring every prefix forms a rooted tree under increasing budgets.
• It employs a bicriterial framework that trades off additive budget slack with multiplicative prize loss, achieving optimal trade-offs on trees and approximations in general graphs.
• Algorithms like DensityGreedy-Tree and capacity-scaling demonstrate practical deployment strategies for phased network connectivity with guaranteed performance at each investment stage.

The Incremental Prize-Collecting Steiner-Tree (Incremental PCST) problem is a variant of the rooted prize-collecting Steiner tree problem where decisions about edge acquisition must be made incrementally, and at each stage any initial segment of the selected edge sequence forms a rooted tree. The problem is governed by a budget constraint that may increase over time, requiring that the solution ordering provide strong guarantees, not just for a single budget but across all budgets simultaneously. The incremental variant is relevant for applications such as phased deployments of network infrastructure, where partial solutions must be high-quality at each investment stage. Recent research addresses the fundamental limits of incremental PCST and develops bicriterial approximation frameworks to quantify algorithmic performance under these constraints (Disser et al., 2024).

1. Formal Problem Definition

Let $G=(V,E)$ be an undirected graph with a designated root vertex $r\in V$, nonnegative vertex prizes $\pi:V\to\mathbb{R}_{\ge 0}$ (with $\pi(r)=0$), and positive edge costs $c:E\to\mathbb{R}_{>0}$. For a subgraph $H\subseteq G$, define the total prize $\pi(H) = \sum_{v \in V_H} \pi(v)$ and the total cost $c(H) = \sum_{e\in E_H} c(e)$.

The static budget-constrained PCST, for budget $B \ge 0$, seeks to maximize: $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$ where $r\in V$0 denotes an (arbitrary) optimal solution.

The incremental PCST requires precommitment to an ordering $r\in V$1 of a subset of $r\in V$2 such that any prefix forms a rooted tree. For realized budget $r\in V$3, let $r\in V$4, and let $r\in V$5 denote the tree of these edges. The solution quality is measured by comparing $r\in V$6 to $r\in V$7, i.e., the incremental algorithm is evaluated at each budget level.

The objective can also be framed as minimizing net cost $r\in V$8, or equivalently maximizing $r\in V$9.

2. Bicriterial Approximation Framework

Unlike the static setting, no ordering can guarantee simultaneous optimality for all possible budgets. Thus, the bicriterial approximation framework permits both additive budget slack and multiplicative prize loss.

An incremental algorithm is $\pi:V\to\mathbb{R}_{\ge 0}$0-competitive if for every $\pi:V\to\mathbb{R}_{\ge 0}$1,

$\pi:V\to\mathbb{R}_{\ge 0}$2

meaning that with additional budget $\pi:V\to\mathbb{R}_{\ge 0}$3, the collected prize is at least $\pi:V\to\mathbb{R}_{\ge 0}$4 times that of the static optimum with budget $\pi:V\to\mathbb{R}_{\ge 0}$5. Neither $\pi:V\to\mathbb{R}_{\ge 0}$6 nor $\pi:V\to\mathbb{R}_{\ge 0}$7 are generally achievable simultaneously for all budgets.

3. Algorithmic Results on Trees

Define, for each $\pi:V\to\mathbb{R}_{\ge 0}$8, $\pi:V\to\mathbb{R}_{\ge 0}$9 is an $\pi(r)=0$0–$\pi(r)=0$1 path$\pi(r)=0$2, and the root's eccentricity as $\pi(r)=0$3. The following algorithms and bounds apply when $\pi(r)=0$4 is a tree:

3.1 Density-Greedy Algorithm

The density of a rooted subtree $\pi(r)=0$5 is $\pi(r)=0$6, with $\pi(r)=0$7. A min-max subtree is an inclusion-minimal rooted subtree with maximal density. On trees, min-max subtrees are edge-disjoint, each incident to the root.

Algorithm DensityGreedy-Tree:

Initiate $\pi(r)=0$8 and an empty ordering $\pi(r)=0$9. While there is remaining prize, operate on the contracted graph $c:E\to\mathbb{R}_{>0}$0, finding a min-max subtree $c:E\to\mathbb{R}_{>0}$1. Append its root-incident edge to $c:E\to\mathbb{R}_{>0}$2, and extend $c:E\to\mathbb{R}_{>0}$3. Repeat until all prizes are collected. The final order $c:E\to\mathbb{R}_{>0}$4 is the incremental sequence.

3.2 Tightness and Guarantees

The DensityGreedy-Tree algorithm is $c:E\to\mathbb{R}_{>0}$5-competitive and runs in polynomial time. That is, with additive slack equal to the root's eccentricity, it is strictly prize-optimal relative to the static solution. No algorithm can achieve $c:E\to\mathbb{R}_{>0}$6 with $c:E\to\mathbb{R}_{>0}$7 and $c:E\to\mathbb{R}_{>0}$8, or $c:E\to\mathbb{R}_{>0}$9 and $H\subseteq G$0. The tightness is witnessed by constructing a two-edge example with eccentricity $H\subseteq G$1.

4. Extension to General Graphs

Let $H\subseteq G$2 be the maximal cost of any vertex-disjoint $H\subseteq G$3–$H\subseteq G$4 path; for trees, $H\subseteq G$5.

4.1 Adapted Density-Greedy and its Limits

The density-greedy approach extends by contracting the current solution, finding a min-max subtree in $H\subseteq G$6, and "lifting" this solution to $H\subseteq G$7 (possibly disconnected). The tree density-greedy algorithm is then applied to each component, appending their edge orderings.

This adapted approach is $H\subseteq G$8-competitive, i.e., with additive slack $H\subseteq G$9, the collected prize is within a factor 2 of the best static solution. No $\pi(H) = \sum_{v \in V_H} \pi(v)$0-competitive guarantee is achievable for any fixed $\pi(H) = \sum_{v \in V_H} \pi(v)$1 in general graphs.

4.2 Approximation and Polynomial-Time Guarantees

Finding an exact maximum-density subtree in general graphs is NP-hard. Employing a $\pi(H) = \sum_{v \in V_H} \pi(v)$2-approximation by Hermans–Lützendörfer–Monnot and a merge-and-swap process, a polynomial-time $\pi(H) = \sum_{v \in V_H} \pi(v)$3-competitive algorithm is achieved, where the multiplicative loss increases to $\pi(H) = \sum_{v \in V_H} \pi(v)$4.

The following table summarizes the tightest known bicriterial guarantees:

Algorithm/Setting	Additive Slack ($\pi(H) = \sum_{v \in V_H} \pi(v)$5)	Multiplicative Loss ($\pi(H) = \sum_{v \in V_H} \pi(v)$6)	Polynomial Time?
DensityGreedy-Tree	$\pi(H) = \sum_{v \in V_H} \pi(v)$7	$\pi(H) = \sum_{v \in V_H} \pi(v)$8	Yes
Adapted Density-Greedy	$\pi(H) = \sum_{v \in V_H} \pi(v)$9	$c(H) = \sum_{e\in E_H} c(e)$0	No (exact)
Hermans–Lützendörfer–Monnot-based	$c(H) = \sum_{e\in E_H} c(e)$1	$c(H) = \sum_{e\in E_H} c(e)$2	Yes

5. Capacity-Scaling Algorithm

The capacity-scaling algorithm reduces dependence on $c(H) = \sum_{e\in E_H} c(e)$3 but allows higher constant multiplicative loss in collected prize, parameterized by $c(H) = \sum_{e\in E_H} c(e)$4.

At each phase $c(H) = \sum_{e\in E_H} c(e)$5, solve (exactly or via static approximation) the PCST for budget $c(H) = \sum_{e\in E_H} c(e)$6, forming a tree $c(H) = \sum_{e\in E_H} c(e)$7. The density-greedy method determines an incremental order within $c(H) = \sum_{e\in E_H} c(e)$8, skipping duplicate or cycle-forming edges. This continues until all prize-vertices are connected.

For any fixed $c(H) = \sum_{e\in E_H} c(e)$9, the algorithm is

$B \ge 0$0

For $B \ge 0$1, this is a $B \ge 0$2-competitive guarantee. As $B \ge 0$3, the multiplicative loss approaches $B \ge 0$4.

This approach leverages a combinatorial lemma: any rooted tree $B \ge 0$5 admits a rooted forest $B \ge 0$6 of at most $B \ge 0$7 components with $B \ge 0$8 and $B \ge 0$9. Applied to the PCST optimum, this justifies the stated bicriterial bounds.

Solving the static PCST exactly is NP-hard in non-tree graphs. Using any static approximation in the base step yields bicriterial results combining both additive and multiplicative losses. Whether polynomial-time algorithms with small constants are achievable remains open.

6. Practical Implications and Open Directions

Bicriterial approximation establishes that, for incremental PCST, near-optimality can be obtained at every budget, at the cost of an $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$0 or $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$1 budget overhead and a constant-factor loss in collected prize. On trees, the $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$2 tradeoff is optimal. For general graphs, $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$3 is achievable (non-polynomially), and $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$4 is available in polynomial time.

Capacity-scaling algorithms connect incrementally ordered solutions in PCST to methods for incremental submodular maximization, though the PCST objective is not submodular, marking a substantial technical distinction.

Key open questions include the possibility of polytime $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$5-competitive incremental algorithms in general graphs, lower bounds on the achievable multiplicative loss $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$6 relative to budget slack $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$7, $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$8, and $\max \left\{ \pi(T) : T \subseteq G,\ T \text{ tree containing } r,\ c(T)\le B \right\}, \tag{P%%%%9%%%%}$9, and extensions to Steiner-forest or variants of the PCST with penalties.

Typical applications arise in the phased deployment of network infrastructure such as fiber or pipelines, where it is essential to obtain near-optimal partial connectivity in early stages while ensuring the overall investment leads to an overall efficient solution (Disser et al., 2024).