Pinwheel Scheduling

Updated 6 July 2026

Pinwheel Scheduling is a scheduling problem where each task with period a_i must appear at least once in every a_i consecutive days, using the density D(A)=∑1/a_i as a feasibility criterion.
The theory establishes sharp feasibility thresholds—like the 5/6 bound—and utilizes state-space models, folding techniques, and computer-assisted proofs to analyze complex scheduling instances.
Variants extend to real periods, matroidal set systems, and optimization problems such as Bamboo Garden Trimming, linking theoretical insights with practical approximation algorithms.

Pinwheel scheduling is a perpetual single-machine scheduling problem in which task $i$ , with period $a_i$ , must be executed at least once in every $a_i$ consecutive time steps. In its standard form, an instance is a finite multiset or ordered tuple of positive integers, one task is scheduled per day, and the central analytical quantity is the density $D(A)=\sum_i 1/a_i$ . The subject has developed around three main themes: exact feasibility criteria and extremal thresholds, finite-state and complexity-theoretic structure, and a network of variants including covering, finite-horizon, real-period, combinatorial, and optimization forms. Two landmarks now anchor the field: the proof that every integer-period instance with density at most $5/6$ is schedulable, settling the Chan–Chin conjecture, and the first NP-hardness theorem for the classical decision problem (Kawamura, 25 Jun 2026, Kleinberg et al., 15 Apr 2026).

1. Classical formulation and state-space view

A classical pinwheel scheduling instance is given by positive integers $A=(a_1,\ldots,a_k)$ , often arranged in nondecreasing order. A schedule is an infinite sequence over $[k]=\{1,\dots,k\}$ , or equivalently a function $S:\mathbb Z\to [k]$ , with exactly one task executed per day. Validity requires that task $i$ appear at least once in every contiguous block of $a_i$ days; in the formulation used by several papers, for every task $a_i$ 0 and every starting day $a_i$ 1, there exists $a_i$ 2 such that $a_i$ 3 (Gąsieniec et al., 2021).

The standard density is

$a_i$ 4

This is a necessary load condition: if $a_i$ 5, then no schedule exists, because at most one task can be executed per day. The condition is not sufficient. Unschedulable families with density at or below $a_i$ 6 are fundamental to the theory; for example, $a_i$ 7 is unschedulable for all $a_i$ 8, and $a_i$ 9 has density $a_i$ 0 but is unschedulable (Kawamura, 25 Jun 2026).

A standard exact model uses a finite directed state graph. One representation stores, at time $a_i$ 1, the vector $a_i$ 2, where $a_i$ 3 is the number of days since task $a_i$ 4 was last executed, with valid states satisfying $a_i$ 5. Another equivalent representation stores remaining waiting times. In either form, an instance is schedulable if and only if the corresponding state graph contains a directed cycle reachable from the initial state; consequently, if an instance is schedulable, then it has a periodic schedule (Gąsieniec et al., 2021).

The literature also uses the name Windows Scheduling for closely related formulations. In the inexact-period version, the time between consecutive executions of job $a_i$ 6 is at most $a_i$ 7; in the exact-period version, it must be exactly $a_i$ 8. For one machine and unit-length jobs, this exact-period model admits a number-theoretic reformulation via Partial Coding (Jacobs et al., 2014).

2. Density theory and sharp schedulability thresholds

Density is the organizing invariant of the classical theory. The long-standing Chan–Chin conjecture asserted that every integer-period instance with density at most $a_i$ 9 is schedulable. This is now a theorem: $D(A)=\sum_i 1/a_i$ 0 The bound is best possible because the family $D(A)=\sum_i 1/a_i$ 1 is unschedulable for all $D(A)=\sum_i 1/a_i$ 2, and its density approaches $D(A)=\sum_i 1/a_i$ 3 as $D(A)=\sum_i 1/a_i$ 4 (Kawamura, 25 Jun 2026).

The proof of the $D(A)=\sum_i 1/a_i$ 5 theorem uses a fractional-period generalization together with weakening, splitting, folding, rounding, and exhaustive finite verification. In that framework, a task with real period $D(A)=\sum_i 1/a_i$ 6 must appear at least $D(A)=\sum_i 1/a_i$ 7 times in any $D(A)=\sum_i 1/a_i$ 8-day interval, equivalently at least $D(A)=\sum_i 1/a_i$ 9 times in every interval $5/6$0 for integer $5/6$1. Folding reduces arbitrary instances to bounded-period families, and a computer-assisted search at threshold $5/6$2 verifies the remaining finite set (Kawamura, 25 Jun 2026).

Before the full proof, the $5/6$3-conjecture had already been attacked through structural computation. The notion of a Pareto surface formalizes the finite set of minimal schedulable instances under dominance for a fixed number of tasks. Using Pareto surfaces and a trie-based search with schedulability pruning, one work confirmed the $5/6$4-conjecture for all instances with at most $5/6$5 tasks and computed the full Pareto surface for $5/6$6, obtaining a list of $5/6$7 periodic schedules that solves every schedulable instance with at most $5/6$8 tasks (Gąsieniec et al., 2021).

The universal threshold $5/6$9 is not the end of density theory. When all periods are bounded below by a larger minimum $A=(a_1,\ldots,a_k)$ 0, stronger guarantees can hold. Defining $A=(a_1,\ldots,a_k)$ 1 as the largest real number such that every instance with all periods at least $A=(a_1,\ldots,a_k)$ 2 and density at most $A=(a_1,\ldots,a_k)$ 3 is schedulable, the first strict improvement over $A=(a_1,\ldots,a_k)$ 4 occurs at $A=(a_1,\ldots,a_k)$ 5: $A=(a_1,\ldots,a_k)$ 6 This was proved by combining fold-based certification with an unfolding operation and a fast heuristic-based solver, showing that once all periods are at least $A=(a_1,\ldots,a_k)$ 7, the guaranteed schedulable region is strictly larger than the universal one (Mishra et al., 25 Aug 2025).

The same density program also yields special-case structure. A byproduct of the fractional-period framework is a simple proof that every instance with at most two distinct period values and density at most $A=(a_1,\ldots,a_k)$ 8 is schedulable; the proof uses a circle-of-circumference- $A=(a_1,\ldots,a_k)$ 9 construction equivalent to complementary Beatty-type scheduling (Kawamura, 25 Jun 2026).

3. Complexity landscape and algorithmic structure

For many years, the exact complexity status of the classical problem remained unsettled. A 2025 finite-variant paper still recorded as open whether classical Pinwheel Scheduling is NP-hard when the deadlines are explicitly listed, whether it lies in NP, and therefore whether it is PSPACE-complete as long conjectured (Kanellopoulos et al., 15 Jul 2025). The first of these questions has since been resolved: Pinwheel Scheduling is NP-hard (Kleinberg et al., 15 Apr 2026).

The NP-hardness proof proceeds by refining an older reduction for Exact Pinwheel Scheduling and padding the constructed instance so that the output density is exactly $[k]=\{1,\dots,k\}$ 0. This density- $[k]=\{1,\dots,k\}$ 1 structure is important both for the reduction itself and for the transfer of hardness to related problems. The same work shows that Dense Pinwheel Scheduling is NP-complete, proves NP-hardness for pinwheel covering, bamboo garden trimming, windows scheduling, recurrent scheduling, and the constant gap problem, and at the same time develops a PTAS, indeed an EPTAS, for the approximate problem: for every fixed $[k]=\{1,\dots,k\}$ 2, there is a polynomial-time algorithm that either declares the instance unschedulable or certifies schedulability of the scaled instance $[k]=\{1,\dots,k\}$ 3 (Kleinberg et al., 15 Apr 2026).

This yields a notable distinction between exact and approximate computation. The exact decision problem is NP-hard, but the approximate feasibility problem with uniform multiplicative slack admits arbitrarily fine polynomial-time approximation. The PTAS relies on a decomposition into big, medium, and small tasks, enumeration of periodic schedules for the big tasks, use of holidays to accommodate medium tasks, and density arguments for sufficiently large periods (Kleinberg et al., 15 Apr 2026).

A different complexity perspective comes from Windows Scheduling. For the exact-period one-machine problem with unit-length jobs, Partial Coding and exact-period Windows Scheduling are polynomial-time interreducible. From this viewpoint, the single-machine exact-period problem does not admit a pseudo-polynomial-time algorithm unless SAT can be solved by a randomized method in expected time $[k]=\{1,\dots,k\}$ 4, and the same lower bound extends to the inexact-period version (Jacobs et al., 2014). These results do not by themselves settle the exact classical complexity question, but they explain why the scheduling problem had resisted standard algorithmic classification.

The field also contains a substantial tradition of computer-assisted exact schedulability tests. In both the $[k]=\{1,\dots,k\}$ 5-density proof and the covering-threshold proof, finite bounded-period families are certified by exhaustive state-graph search augmented with symmetry reduction, folding reductions, and witness reuse (Kawamura, 25 Jun 2026, Kawamura et al., 8 Oct 2025).

4. Generalizations beyond integer periods and single-task actions

One line of extension replaces integer periods by real periods. In the real-period model, task $[k]=\{1,\dots,k\}$ 6 with real period $[k]=\{1,\dots,k\}$ 7 must appear at least $[k]=\{1,\dots,k\}$ 8 times in any consecutive $[k]=\{1,\dots,k\}$ 9 days for every positive integer $S:\mathbb Z\to [k]$ 0. This reduces to the usual definition when $S:\mathbb Z\to [k]$ 1 is integral and captures genuinely stronger frequency constraints than a naive “once every $S:\mathbb Z\to [k]$ 2 days” interpretation. For this model, it has been proved that any real-valued instance with three distinct period values and density at most $S:\mathbb Z\to [k]$ 3 admits a valid schedule, and a corresponding real-period analogue of the Chan–Chin conjecture has been proposed (Fujiwara et al., 28 Oct 2025).

Another extension replaces the “one task per day” machine model by arbitrary set systems. In Combinatorial Pinwheel Scheduling (CPS), an instance is $S:\mathbb Z\to [k]$ 4, where $S:\mathbb Z\to [k]$ 5 is a set system and each time step schedules an independent set $S:\mathbb Z\to [k]$ 6. The requirement is that each element $S:\mathbb Z\to [k]$ 7 appear at least once in every window of $S:\mathbb Z\to [k]$ 8 consecutive time steps. In this setting, density becomes the minimum $S:\mathbb Z\to [k]$ 9 for which there exist coefficients $i$ 0 whose total sum is $i$ 1 and whose aggregate support gives each element $i$ 2 frequency at least $i$ 3. For matroid constraints, density at most $i$ 4 is sufficient for schedulability; for arbitrary set systems, no constant density threshold independent of $i$ 5 is possible, and the best guarantee scales as $i$ 6 (Mendoza-Cadena et al., 12 Feb 2026).

The set-system viewpoint connects pinwheel scheduling to combinatorial bamboo garden trimming and to matroid algorithms. For uniform and partition matroids, polynomial-time height- $i$ 7 schedules are available; for graphic and laminar matroids, polynomial-time height- $i$ 8 schedules are obtained. The uniform-matroid algorithm is the Fuse–Unfuse schedule (Mendoza-Cadena et al., 12 Feb 2026).

A further application-driven generalization appears in multi-hop wireless networks with hard deadlines. There, end-to-end delay bounds are expressed in terms of per-link maximum inter-scheduling times, and the resulting joint feasibility problem is a generalized pinwheel coloring problem under conflict-graph interference constraints. Under total interference, the formulation reduces to ordinary pinwheel scheduling; with general interference it becomes a link-coloring-and-period problem, again NP-hard, for which a decentralized polynomial-time heuristic is proposed (Jones et al., 19 Apr 2026).

5. Finite-horizon and covering variants

The most prominent finite-horizon analogue is the $i$ 9-Visits problem. Given deadlines $a_i$ 0, the task is to construct a schedule of length $a_i$ 1 in which each task appears exactly $a_i$ 2 times and each occurrence is at most $a_i$ 3 positions away from the previous one, or from the beginning of the schedule for the first occurrence. The case $a_i$ 4 is trivial, but $a_i$ 5-Visits is strongly NP-complete. The same line of work established a sharp dichotomy: if all deadlines are distinct, then $a_i$ 6-Visits is solvable in linear time, whereas multiplicities make the problem hard (Kanellopoulos et al., 15 Jul 2025).

Subsequent work strengthened this complexity picture by proving that $a_i$ 7-Visits remains strongly NP-complete even when the maximum multiplicity of the input is $a_i$ 8. It also placed $a_i$ 9-Visits in RP when the number of distinct deadlines per cluster of the discretized sequence is constant, generalized the positive results to a mixed $a_i$ 00-Visits model, and showed that the lower density threshold of $a_i$ 01-Visits is at least

$a_i$ 02

More broadly, the lower density threshold of $a_i$ 03-Visits approaches $a_i$ 04 as $a_i$ 05, linking the finite and infinite theories (Kanellopoulos et al., 17 Apr 2026).

The dual of classical pinwheel scheduling is the covering problem, also called discretized point patrolling. Here one daily task is assigned to agents under the constraint that agent $a_i$ 06 may perform the task at most once in any $a_i$ 07-day interval. Its density uses the same sum $a_i$ 08, but the extremal direction is reversed: the question is how large the density must be to force schedulability. The optimal covering threshold is

$a_i$ 09

and every covering instance with density at least $a_i$ 10 is schedulable. The bound is tight because the family

$a_i$ 11

is unschedulable and has density approaching $a_i$ 12 from below (Kawamura et al., 8 Oct 2025).

The covering proof is structurally parallel to the packing proof, but not identical. It uses a real-valued extension, a folding lemma that reduces unschedulable instances to bounded ranges of periods, a finite exhaustive lemma on periods in $a_i$ 13, and an additional reduction that peels off the initial $a_i$ 14-type prefix. A related presentation in the discretized point patrolling literature describes improved fold operations and density-loss accounting that also recover the same optimal threshold (Mishra, 24 Oct 2025).

The optimization counterpart of pinwheel scheduling is Bamboo Garden Trimming (BGT). In the discrete model, bamboo $a_i$ 15 grows at rate $a_i$ 16, one bamboo may be trimmed to height zero per day, and the goal is to minimize the maximum height ever attained. The link to pinwheel scheduling is exact at the level of decision thresholds: if the maximum allowable height is $a_i$ 17, then the BGT instance corresponds to the pinwheel instance $a_i$ 18 (Kawamura, 25 Jun 2026).

This reduction lets pinwheel theory translate into approximation algorithms. Using the $a_i$ 19-density theorem, one paper gives a polynomial-time $a_i$ 20-approximation algorithm for BGT (Kawamura, 25 Jun 2026). A later work improves this to a $a_i$ 21-approximation, again by combining fold-based pinwheel reasoning with lookup tables and special handling of low periods such as $a_i$ 22 and $a_i$ 23 (Mishra, 24 Oct 2025). Earlier, a pinwheel-based algorithm for BGT had already improved the approximation ratio below $a_i$ 24, proving a worst-case bound of $a_i$ 25 and an asymptotic ratio converging to $a_i$ 26 when $a_i$ 27 (Croce, 2020).

Pinwheel scheduling also serves as the star-graph special case of broader graph-theoretic scheduling models. In Polyamorous Scheduling, the tasks are edges of a graph and each day’s feasible action is a matching; the optimization version minimizes the maximum weighted waiting time between consecutive occurrences of the same edge. On a star graph, this becomes classical pinwheel scheduling in the decision form and bamboo garden trimming in the optimization form. The model introduces a generalized density notion, poly density, via an LP dual lower bound, and asks whether a constant feasibility threshold analogous to $a_i$ 28 exists in this graph setting (Gąsieniec et al., 2024).

The broader significance of pinwheel scheduling lies in this capacity to act simultaneously as a sharp extremal theory, a source of hardness reductions, and a reusable scheduling core. Its modern landscape includes an exact density threshold for the classical packing problem, an exact dual threshold for covering, finite-horizon variants with their own threshold behavior, generalizations to real periods and matroidal set systems, and approximation frameworks for optimization problems built on the same recurrent-service constraint (Kawamura, 25 Jun 2026, Kawamura et al., 8 Oct 2025, Kleinberg et al., 15 Apr 2026).