Freeze-Tag Problem (FTP) Overview

• Freeze-Tag Problem is a scheduling and activation challenge where one active robot awakens others, modeled via rooted wake-up trees to minimize makespan.
• It employs spanning tree representations with weighted edges under diverse metrics including Euclidean, Manhattan, and geodesic distances.
• Research spans complexity analysis, approximation algorithms, and variants (online, distributed), impacting robotics and motion planning.

Searching arXiv for recent and foundational Freeze-Tag papers to support the article. arXiv search query: "Freeze-Tag Problem geometric online distributed polygonal domain" The Freeze-Tag Problem (FTP) is a scheduling and activation problem in which one initially awake robot must awaken a set of sleeping robots by moving to their positions; each awakened robot immediately becomes awake and can assist in awakening others (Rajabi-Alni et al., 2024). In standard metric formulations, a schedule can be represented by a rooted wake-up tree whose weighted depth is the makespan, the time when the last robot becomes active (Brunner et al., 2019).

1. Formal model and objective

An FTP instance is commonly written as $(R,r_0)$, where $R$ is a finite set of robot positions in a metric space $(M,d)$, and $r_0$ is the distinguished initially active robot. All other robots are initially frozen. Active robots move at unit speed; when an active robot reaches the exact location of a frozen robot, that robot wakes up instantly and can itself begin moving. The optimization objective is to minimize the makespan, namely the maximum activation time over all robots (Silva et al., 17 Sep 2025).

A convenient formal representation is a rooted wake-up tree. The vertices are the robot positions, the root is the source, and an edge $(r_i,r_j)$ means that some active robot reaches $r_j$ from $r_i$ and activates it. In metric formulations, the edge weight is $d(r_i,r_j)$, and the makespan is the maximum weighted root-to-leaf distance. In the standard tree view used in the literature, the root is rooted at $r_0$, and the schedule is equivalent to finding a spanning binary tree of minimum weighted depth (Brunner et al., 2019).

Geometric formulations often normalize the instance by the radius

$r=\max_{i\ge 1} d(p_0,p_i),$

or, in polygonal domains, by the diameter

$R$0

where $R$1 is a geodesic path inside the domain. These quantities serve both as lower bounds on any feasible makespan and as reference scales for approximation guarantees (Rajabi-Alni et al., 2024).

2. Metric and geometric settings

FTP has been studied under several distance models. In geometric settings, the distinction between Euclidean distance, Manhattan distance, geodesic distance inside obstacles, visibility-based distance, and time-dependent or distributed information models materially changes both the combinatorics of wake-up trees and the achievable guarantees.

Setting	Distance model	Structural feature
General metric FTP	Arbitrary metric $R$2	Wake-up tree with weighted depth
Euclidean / $R$3 FTP	$R$4-norm in $R$5	Robots are points in free space
GFTP	Geodesic distance $R$6 in polygonal domain $R$7	Symmetric metric inside obstacles
VFTP	Geodesic visibility distance $R$8	Directed, generally asymmetric
Online TDFT	Metric plus release times $R$9	Requests are revealed over time
Distributed dFTP	Euclidean plane, local sensing radius $(M,d)$0	Positions initially unknown

In polygonal domains, the domain $(M,d)$1 may contain holes, robots are arbitrary points of $(M,d)$2, and motion must remain inside $(M,d)$3. Two complete graph structures are central there: the geodesic graph $(M,d)$4, with edges labeled by shortest geodesic paths, and the directed geodesic visibility graph $(M,d)$5, where an edge from $(M,d)$6 to $(M,d)$7 records the shortest path that $(M,d)$8 must travel until $(M,d)$9 becomes visible (Rajabi-Alni et al., 2024).

In the online time-dependent model, requests are pairs $r_0$0, where the position $r_0$1 is not known before its release time $r_0$2. In the distributed model, by contrast, there is no advance knowledge of the sleeping positions at all: awake robots only detect robots within Euclidean distance $r_0$3, share a global clock and coordinate system once awake, and must interleave exploration with awakening (Brunner et al., 2019, Gavoille et al., 28 Mar 2025).

3. Complexity landscape

The complexity theory of FTP is broad and strongly negative. On graphs, FTP is NP-hard even on star graphs with the same number of robots at each leaf; for that setting the literature reports a PTAS, a greedy approximation algorithm whose approximation factor has a lower bound of $r_0$4, and a $r_0$5-approximation when leaves may hold different numbers of robots (Rajabi-Alni et al., 2024).

In continuous geometry, hardness persists in low dimensions. FTP is NP-hard in the Euclidean plane under $r_0$6, NP-hard in $r_0$7 under any $r_0$8 distance with $r_0$9, and the long-open planar Manhattan case was resolved by proving that FTP is strongly NP-hard in $(r_i,r_j)$0 with $(r_i,r_j)$1 distance. That $(r_i,r_j)$2 result also yields strong NP-completeness with integer coordinates and NP-completeness on unweighted grid graphs (Silva et al., 17 Sep 2025).

Obstacle geometry does not simplify the exact problem. Inside simple polygons, both the geodesic and visibility versions, GFTP and VFTP, are NP-hard; the later extension to polygonal domains with holes preserves the difficulty and shifts attention to approximation algorithms (Rajabi-Alni et al., 2024). Likewise, 3D Manhattan geometry is NP-complete: a reduction from Monotone-3SAT establishes FTP hardness in $(r_i,r_j)$3, closing the $(r_i,r_j)$4 gap left by earlier $(r_i,r_j)$5, $(r_i,r_j)$6, hardness results (Silva, 2023).

For the online time-dependent variant, complexity is expressed competitively rather than by NP-hardness alone. No online algorithm can achieve competitive ratio better than $(r_i,r_j)$7, and the exact optimal competitive ratio over continuous metric spaces is $(r_i,r_j)$8 (Brunner et al., 2019).

4. Approximation algorithms and quantitative bounds

Approximation theory for FTP is highly geometry-dependent. Euclidean FTP admits both constant-factor approximation and PTAS techniques, and the polygonal-domain literature imports that paradigm through spanners and balanced trees. In polygonal domains with holes, a constant-factor algorithm for GFTP is obtained by adding Steiner robots at reflex and hole vertices, constructing a visibility $(r_i,r_j)$9-spanner of maximum degree at most $r_j$0, and adapting the classical constant-factor awakening strategy to that bounded-degree spanner. The resulting bound is

$r_j$1

hence makespan $r_j$2; the same spanner transfer yields an $r_j$3-approximation for VFTP as well. The same paper also gives PTASs for both GFTP and VFTP via pixel decomposition, one representative robot per pixel, dynamic programming over pseudo-balanced awakening trees, and local expansion inside each pixel (Rajabi-Alni et al., 2024).

In planar $r_j$4 geometry, the wake-up constant is known exactly. If all sleeping robots are within $r_j$5-distance $r_j$6 of the source, a makespan of at most $r_j$7 can always be achieved, and a schedule achieving that bound can be computed in optimal time $r_j$8; both the time bound and the makespan bound are optimal. Through norm comparison, this also implies $r_j$9 in the Euclidean plane (Bonichon et al., 2024).

Subsequent work tightened Euclidean constants substantially. A 2024 geometric study reported $r_i$0, $r_i$1, and therefore $r_i$2 (Alipour et al., 2024). A later improvement reduced the planar Euclidean upper bound to $r_i$3, then summarized the further improvement to $r_i$4, and finally established

$r_i$5

The same line of work records the standard lower bound in $r_i$6 as about $r_i$7, leaving a narrow but nonzero gap between known lower and upper bounds (Alipour et al., 22 Jul 2025).

The algorithmic techniques behind these constants are varied but recurring: crown and cone decompositions in two dimensions, recursive region partition in three dimensions, bounded-degree spanners, pseudo-balanced trees, and local-to-global replacement arguments. In particular, the polygonal PTAS and the Euclidean wake-up-ratio analyses both rely on replacing full robot sets by smaller structured summaries—representatives, spanner neighborhoods, or recursively defined sectors—without losing control of the weighted depth of the wake-up tree (Rajabi-Alni et al., 2024, Alipour et al., 22 Jul 2025).

5. Variants beyond the centralized offline model

A major line of work studies FTP when information is not available in advance. In online time-dependent freeze-tag, robots are revealed only at release times $r_i$8. The optimal online competitive ratio is $r_i$9, achieved by a strategy that, after each new release, sends active robots home, waits until time $d(r_i,r_j)$0 for the currently known prefix, executes an optimal offline schedule for that prefix, and then returns home again. The lower bound matches the upper bound on every continuous metric space (Brunner et al., 2019).

The distributed version changes the problem more radically. In distributed FTP, sleeping positions are unknown, awake robots only detect others within distance $d(r_i,r_j)$1, and exploration cost becomes intrinsic. With $d(r_i,r_j)$2 denoting the largest Euclidean distance from the source to any sleeping robot and $d(r_i,r_j)$3 the connectivity threshold of the robot set, the whole swarm can be awakened in time

$d(r_i,r_j)$4

and this dependence on $d(r_i,r_j)$5 and $d(r_i,r_j)$6 is matched by a lower bound. Under bounded energy, one algorithm uses budget $d(r_i,r_j)$7 per robot and achieves makespan $d(r_i,r_j)$8, while another uses budget $d(r_i,r_j)$9 and achieves

$r_0$0

again with complementary lower bounds in the applicable parameter range (Gavoille et al., 28 Mar 2025).

Other variants constrain initial activation or terminal conditions. The multi-source version $r_0$1-FTP allows more than one initially awake robot; the Euclidean literature summarized in the polygonal-domain paper includes a PTAS for $r_0$2-FTP (Rajabi-Alni et al., 2024). Freeze-Tag-with-Return (FTRP) requires all robots eventually to return to initial positions. In the Euclidean unit disk, the difference between the optimal makespans of FTP and FTRP never exceeds $r_0$3 and is at least $r_0$4 in the worst case. For sleeping robots in convex position, the optimal FTRP makespan is at most $r_0$5, and the problem also admits single-exponential algorithms for general distance functions, with ETH-based optimality in metric spaces and NP-hardness on unweighted graphs (Bonichon et al., 20 Jun 2026).

6. Broader structure, methodology, and open directions

FTP sits at the intersection of geometric approximation, broadcast-tree design, and motion planning. The literature repeatedly recasts the problem as one of constructing shallow rooted trees under geometric or graph-derived edge weights, then exploits whatever structure the ambient space offers: monotone shortest paths in $r_0$6, cone and crown decompositions in Euclidean disks, visibility spanners in polygons, or recursive coverings of cross-polytopes in three dimensions (Bonichon et al., 2024, Alipour et al., 22 Jul 2025, Rajabi-Alni et al., 2024). This methodological convergence is notable because the underlying hardness landscape is already broad: exact optimization is difficult, but the best approximation strategies are highly sensitive to geometry.

Several research directions remain explicitly open. For polygonal domains, open directions include multi-source FTP inside polygons, improving constants such as the factor $r_0$7, and extending techniques to higher-dimensional obstacle environments (Rajabi-Alni et al., 2024). For Euclidean wake-up ratios, the exact value of $r_0$8 remains unknown, and improved constants in three dimensions are still sought (Alipour et al., 22 Jul 2025). In online settings, a natural question is whether polynomial-time online algorithms can approach the optimal ratio $r_0$9 when exact recomputation of offline optima is intractable (Brunner et al., 2019). In distributed settings, open problems include matching optimal makespan with only $r=\max_{i\ge 1} d(p_0,p_i),$0 energy and eliminating the face-to-face communication assumption (Gavoille et al., 28 Mar 2025). For FTRP, the central conjecture is

$r=\max_{i\ge 1} d(p_0,p_i),$1

together with the unresolved Euclidean hardness of the return-constrained problem itself (Bonichon et al., 20 Jun 2026).

Taken together, these results depict FTP as a canonical activation problem whose difficulty is stable across models but whose algorithmic behavior depends delicately on metric structure. Graphs, Euclidean spaces, polygonal domains, release times, local sensing, and return constraints all preserve the core wake-up tree viewpoint, yet each alters the achievable guarantees, the dominant proof techniques, and the set of meaningful open questions.