On-demand Steiner Tree (OST)

• OST is a family of Steiner tree formulations that process terminals at query, arrival, or request time, optimizing low-cost connectivity.
• It spans multiple regimes including online, fully dynamic, exact query, and heterogeneous-demand multicast, each with distinct recourse and preprocessing strategies.
• Advanced methods use prediction, limited recourse, and geometric preprocessing to improve performance, reduce computational overhead, and enhance solution quality.

Searching arXiv for the cited Steiner tree and OST papers to ground the article in recent and canonical sources. On-demand Steiner Tree (OST) denotes a family of Steiner-tree formulations in which required terminals or service demands are processed at query time, arrival time, or request time rather than being handled only as a single static offline instance. In current arXiv literature, the phrase covers online Steiner tree with irrevocable terminal arrivals, exact query-oriented Steiner computation on large static graphs after preprocessing, goal-oriented exact search over Steiner dynamic-programming states, and a heterogeneous-demand multicast formulation for 6G multimedia delivery (Xu et al., 2021, Wei-Kleiner, 2013, Hougardy et al., 2014, Wang et al., 7 Jul 2025). Across these settings, the common core is the construction or maintenance of a low-cost tree connecting required vertices; the principal differences concern when the demand set becomes known, what recourse is permitted, whether preprocessing is available, and whether edge cost reflects only topology or also carried flow.

1. Formal scope and problem regimes

The underlying reference problem is the Steiner tree problem: given a weighted graph or a planar terminal set, construct a minimum-cost tree connecting a specified terminal set. OST-related work specializes this template in several directions. In graph-theoretic exact-query settings, the input is a weighted undirected graph $G=(V,E)$ with positive edge weights and a terminal set $S\subseteq V$, and the goal is an exact minimum-weight tree spanning $S$ (Wei-Kleiner, 2013). In online settings, terminals $t_1,t_2,\dots,t_k$ arrive one at a time and edges, once bought, are irrevocable; the objective is to minimize total purchased cost while preserving connectivity among seen terminals (Xu et al., 2021). In fully dynamic settings, the graph is fixed but the terminal set changes through additions and removals, and the task is to maintain an approximate Steiner tree efficiently under updates (Łącki et al., 2013). In the heterogeneous-demand multicast formulation, there is a single source, multiple destinations, destination-specific demands $x^j$, and the objective is to minimize the weighted sum of transmitted flow (Wang et al., 7 Jul 2025).

Regime	Information pattern	Objective
Online OST	terminals arrive online	minimize total purchased edge cost
Dynamic Steiner maintenance	terminal additions and removals	maintain an approximate Steiner tree quickly
Exact on-demand query processing	static graph, repeated small-terminal queries	return an exact minimum-weight Steiner tree
Heterogeneous-demand OST	one source, multiple destinations with demands $x^j$	minimize total weighted flow cost

A common source of ambiguity is that OST is not tied to a single standardized mathematical model. In the online literature, it refers to adversarial or prediction-augmented arrival sequences; in exact-query literature, it denotes demand-driven computation on a large static graph; in the 6G multicast literature, it denotes a specific dynamic-programming algorithm for heterogeneous outflow requirements. What remains invariant is the exploitation of shared connectivity: unlike independent shortest paths, all OST variants derive value from reusing common prefixes or separators.

2. Online OST and prediction-augmented algorithms

In the online Steiner tree problem on undirected graphs, the main analysis assumes metric completion: “we assume that $G$ is a complete graph in metric space.” A terminal set $T\subseteq V$ arrives online, and when a terminal appears, the algorithm must immediately connect it to previously seen terminals by buying irrevocable edges. In directed graphs, a root $r$ is fixed and every terminal must obtain a directed path to $r$ in the bought subgraph (Xu et al., 2021). The classical worst-case guarantees are sharp: the greedy algorithm is $S\subseteq V$0-competitive in undirected graphs and $S\subseteq V$1-competitive in directed graphs, and these are best possible under worst-case analysis.

The learning-augmented model enriches this by supplying, at time $S\subseteq V$2, a predicted terminal set $S\subseteq V$3. The prediction error is the number of incorrectly predicted terminals; under the equal-size assumption $S\subseteq V$4, the paper uses

$S\subseteq V$5

and when equal size is dropped, the appendix redefines error as

$S\subseteq V$6

This model predicts terminal identity rather than arrival order. The undirected warm-up algorithm OAPT computes $S\subseteq V$7 and routes later arrivals either by a cheapest connecting edge or along the predicted MST; its competitive ratio is $S\subseteq V$8. The improved algorithm IOAPT modifies the prediction-following step by buying only a subpath $S\subseteq V$9 with cost in $S$0, and achieves $S$1. The lower bound is matching up to constants: no online algorithm can do better than $S$2 for online undirected Steiner tree with predicted terminals. In the directed case, the algorithm maintains a guessed threshold $S$3, a predicted subset $S$4, and $S$5; it attains competitive ratio $S$6, and a lower bound $S$7 shows the linear dependence on prediction error is essentially unavoidable (Xu et al., 2021).

The structural intuition is that predictions reduce uncertainty about which vertices will matter later, allowing the algorithm to pre-build shared infrastructure around likely terminals. The empirical results are consistent with that interpretation. On random graphs and Bay Area DIMACS road graphs, OAPT and IOAPT outperform the online greedy baseline once about $S$8 of the predictions are correct. When predictions are learned from samples, difficult distributions such as uniformly random terminals cause the learner to predict almost nothing, which preserves robustness; on learnable distributions such as the two-class random graph distribution or clustered road-graph distributions, a small number of training instances yields useful predictions, and after tens of training instances the reported ratios become never worse than $S$9 on difficult distributions (Xu et al., 2021).

3. Recourse and fully dynamic maintenance

A distinct line of OST-related work relaxes irrevocability by allowing limited recourse. In the strengthened online model, each arrival still requires adding one edge that connects the new vertex, but the algorithm may also swap old edges. The central result is that one swap per arrival suffices to maintain a constant-competitive tree in the metric Steiner tree problem (Gu et al., 2013). More generally, for any $t_1,t_2,\dots,t_k$0, there is an online $t_1,t_2,\dots,t_k$1-competitive algorithm that performs at most one swap upon each arrival and at most $t_1,t_2,\dots,t_k$2 swaps per arrival amortized. This sharply separates the plain irrevocable model, where greedy is only $t_1,t_2,\dots,t_k$3-competitive, from the limited-recourse model, where constant competitiveness becomes attainable.

The technical framework is primal-dual and rank-based. A clustering process defines vertex ranks $t_1,t_2,\dots,t_k$4 by moat growth in phases, with root $t_1,t_2,\dots,t_k$5 having rank $t_1,t_2,\dots,t_k$6. The dual lower bound is expressed through the rank-weight function, which satisfies

$t_1,t_2,\dots,t_k$7

while valid trees with respect to an admissible function $t_1,t_2,\dots,t_k$8 obey

$t_1,t_2,\dots,t_k$9

To control recourse, the algorithm maintains virtual ranks $x^j$0 that lag behind actual ranks but change in only $x^j$1 coordinates per step in the constant-budget construction; the one-swap version then follows by a sparsified charging argument (Gu et al., 2013).

The fully dynamic viewpoint is different. Here the graph is fixed and the terminal set changes by insertions, deletions, or both; the objective is fast update time rather than a bound on the number of structural changes. A central reduction is to the metric closure $x^j$2: maintaining an approximate MST on active terminals in $x^j$3 yields a factor-2 approximation to Steiner tree in the original graph. The framework uses dynamic vertex-color distance oracles and dynamic minimum spanning forest maintenance. In general graphs it maintains a fully dynamic $x^j$4-approximate Steiner tree in $x^j$5 amortized update time, and in planar graphs it achieves $x^j$6 with the same asymptotic update time; faster incremental and decremental variants are also given, as well as a polylogarithmic-time $x^j$7-approximate algorithm for planar graphs when a higher approximation factor is allowed (Łącki et al., 2013). The paper explicitly distinguishes this from the classical online viewpoint: online algorithms minimize structural changes after each update, whereas dynamic algorithms minimize the time needed to process each update.

4. Exact on-demand computation on static graphs

When the graph is large and static but queries involve a bounded number of terminals, OST takes the form of exact on-demand query answering with heavy offline preprocessing. The tree-decomposition-based algorithm STEIN I is designed for precisely this regime. It assumes a nice tree decomposition whose internal nodes are leaf, introduce vertex, forget/remove vertex, or join/merge nodes. The essential separator fact is that every bag on the path between two vertices in a tree decomposition is a vertex separator for those vertices. Using this, the algorithm precomputes Steiner-tree sets $x^j$8 for each bag $x^j$9 and then answers a query by locating terminal-induced subtree roots, finding their lowest common ancestor, and traversing bottom-up while combining states through vertex removal, vertex insertion, and merge operations (Wei-Kleiner, 2013).

The method is exact and explicitly intended for repeated queries on large graphs. It is motivated by the fact that Dreyfus–Wagner has complexity

$x^j$0

whose cubic dependence on $x^j$1 is prohibitive even when $x^j$2 is constant. STEIN I shifts the dependence away from $x^j$3 at query time. Its query complexity is

$x^j$4

where $x^j$5 is the height of the tree decomposition and $x^j$6 its width, while index storage is

$x^j$7

The central separator-based recurrence, STVS, minimizes over separator vertices $x^j$8 and bipartitions $x^j$9. Its raw complexity is $G$0, but the decomposition supplies small separators systematically and makes the relevant sub-Steiner trees precomputable (Wei-Kleiner, 2013).

A complementary exact approach is goal-oriented rather than index-oriented. “Dijkstra meets Steiner” replaces exhaustive subset-order dynamic programming by a label-setting search over states $G$1, where each label represents the best known cost for a partial Steiner tree connecting $G$2. The algorithm repeatedly finalizes the non-permanent label minimizing

$G$3

where $G$4 is a valid lower bound on the future cost of connecting the remaining terminals. This is the Steiner analogue of Dijkstra or $G$5 with feasible potentials. The worst-case runtime is

$G$6

matching the best known asymptotic range when $G$7 is chosen from the paper’s cheap lower bounds, but practical behavior is substantially improved by guided label expansion and pruning (Hougardy et al., 2014). On some large VLSI-derived instances, reported runtimes improve previous best values by orders of magnitude, and the method solves higher-dimensional rectilinear Steiner instances whose Hanan grids contain up to several millions of edges (Hougardy et al., 2014).

These exact-query methods clarify a frequent misconception: “on demand” does not imply recomputation from scratch on the full graph. In one design, preprocessing compresses the search into separators and bag states; in the other, future-cost guidance ensures that only promising states are explored.

5. Heterogeneous-demand multicast OST

A recent and more specialized use of OST appears in 6G multimedia delivery, where the problem is no longer merely to connect destinations but to satisfy heterogeneous outflow requirements at minimum transmission cost. The network is an undirected weighted graph $G$8 with one source, multiple destinations, and routing nodes. Each destination $G$9 carries a demand $T\subseteq V$0, interpreted as a video-resolution ratio such as $T\subseteq V$1 for $T\subseteq V$2. The objective is the weighted sum of transmitted flow, subject to flow-conservation-like constraints at intermediate nodes, sufficient source injection to satisfy the largest downstream demand, per-destination demand satisfaction, nonnegativity, and feasibility on existing edges (Wang et al., 7 Jul 2025).

The first structural theorem states that if $T\subseteq V$3, the used links of an optimal solution form a tree rooted at the source, with destinations as leaves and flow directed from root toward leaves. The proof eliminates disconnected components, cycles, multiple disjoint paths between the same nodes, and useless non-destination leaves, each time reducing cost without violating feasibility. A second key lemma states that if a link is shared by multiple destination flows, the flow on that link equals the maximum demand among those destinations. This max-aggregation rule is what distinguishes the model from classical Steiner tree, where terminals are ordinarily binary endpoints without differentiated rate requirements (Wang et al., 7 Jul 2025).

The resulting OST algorithm is a two-stage dynamic program over states $T\subseteq V$4, the minimum total flow cost needed to connect node $T\subseteq V$5 to a destination subset $T\subseteq V$6. The boundary conditions are

$T\subseteq V$7

Stage 1 performs subset partitioning: $T\subseteq V$8 reflecting the fact that a shared branch must carry the larger of the merged downstream demands. Stage 2 performs edge extension: $T\subseteq V$9 which is a shortest-path-like recurrence with demand-weighted edge cost. The optimality proof is by induction on the number of destinations and shows that the dynamic program computes the true minimum flow cost $r$0 exactly (Wang et al., 7 Jul 2025).

The method is exact but exponential in the number of destinations: $r$1 time and

$r$2

space. Experiments use random link costs $r$3 and demands sampled from $r$4 with equal probability, and compare against MST, Dijkstra’s Algorithm, Genetic Algorithm, Ant Colony Optimization, and Bee Colony Optimization. The reported primary outcome is that OST reduces total network flow by over $r$5 compared to state-of-the-art methods while satisfying on-demand QoS fulfillment (Wang et al., 7 Jul 2025). In this formulation, “on-demand” does not mean online arrival of terminals; it means destination-specific service rates that must be honored jointly with shared multicast routing.

6. Geometric preprocessing and Euclidean 1-Steiner acceleration

OST-related methodology also appears in Euclidean Steiner optimization, where the emphasis is on geometric restriction of the candidate space rather than on online arrivals or repeated graph queries. For the Euclidean $r$6-Steiner problem, overlaid oriented Voronoi diagrams (OOVDs) encode local directional nearest-neighbor information that sharply constrains where an optimal Steiner point can be placed. The plane is partitioned into six cones of angle $r$7; six oriented Voronoi diagrams are built and then overlaid, together with the classical Voronoi diagram, to form the refined OOVD (Payne et al., 2020).

Each OOVD face stores a 7-entry vector: six entries for the closest terminal in each cone direction and one for the globally closest terminal. If $r$8 is such a vector and $r$9 its face, then any Steiner point $r$0 can only have neighbors among the terminals listed in $r$1. Classical geometric facts then reduce the candidate neighbor sets drastically. A Steiner point in an optimal Euclidean $r$2-Steiner tree has degree at most $r$3. In the degree-$r$4 case, the incident edges meet at $r$5; because the cones have angle $r$6, only the two 3-terminal subsets $r$7 and $r$8 are feasible. In the degree-$r$9 case, the neighbors are in convex position and only three 4-terminal subsets are feasible. Consequently each OOVD face yields at most $S\subseteq V$00 buckets before pruning, and at most $S\subseteq V$01 once the Steiner point is required to be adjacent to its closest terminal $S\subseteq V$02 (Payne et al., 2020).

This geometric preprocessing changes the effective complexity profile of the $S\subseteq V$03-Steiner search. The naive method tests all $S\subseteq V$04 candidate neighbor sets and has asymptotic complexity $S\subseteq V$05, assuming MST construction in $S\subseteq V$06. The OOVD-based method processes the refined face data, constructs an MST, precomputes a table of longest edges on tree paths, and then evaluates only the reduced bucket list; the stated overall complexity is $S\subseteq V$07 (Payne et al., 2020). The implementation uses CGAL upper-envelope computation and planar-map overlay, and exactness is restored by representing coordinates in the quadratic extension field $S\subseteq V$08, avoiding spurious thin faces caused by rational-rounding issues.

The empirical findings are notable. On uniformly random points in a $S\subseteq V$09 integer grid, with $S\subseteq V$10 and $S\subseteq V$11 instances each, the number of refined OOVD faces appeared essentially linear in $S\subseteq V$12, and the refined OOVD had a little under $S\subseteq V$13 times as many cells as terminals. For a $S\subseteq V$14-point input, the run completed in just over $S\subseteq V$15 minutes, produced $S\subseteq V$16 regions, and used about $S\subseteq V$17 GB RAM at peak. Processing the OOVD data reduced $S\subseteq V$18-Steiner construction time by roughly a factor of $S\subseteq V$19, directly explained by the bucket reduction from $S\subseteq V$20 to at most $S\subseteq V$21 per face (Payne et al., 2020). This suggests a geometric analogue of OST’s broader design principle: invest in structure that restricts candidate subproblems to those that can actually contribute to an optimum.