---
title: 'Social Golfer Problem: Designs & Algorithms'
url: https://www.emergentmind.com/topics/social-golfer-problem-sgp
type: topic
---

# Social Golfer Problem: Designs & Algorithms

Searching arXiv for recent and foundational papers on the Social Golfer Problem.
The Social Golfer Problem (SGP) is the problem of scheduling \(q=g\cdot p\) golfers over \(w\) weeks so that, in each week, the golfers are partitioned into \(g\) groups of size \(p\), every golfer plays exactly once per week, and no pair of golfers plays together more than once over the whole schedule [1406.7196]. In combinatorial-design terminology, the same problem is written as \(\text{SGP}(v,k,r)\), where \(v=k\cdot n\) points are arranged into \(r\) rounds of \(n\) non-intersecting blocks of size \(k\), again with the condition that any two points are assigned to the same block at most once [2507.23376]. Recent work places this classical offline design problem alongside a distinct greedy or online variant in which rounds are chosen one by one with no preplanning and no revision of earlier rounds, yielding a sharp worst-case lower bound on how many rounds can always be guaranteed [2007.10704].

## 1. Formal definitions and equivalent representations

Three parameterizations recur in the literature. In the scheduling formulation, an instance is given by a triple \(g\)-\(p\)-\(w\), with \(q=g\cdot p\) total golfers, \(g\) groups per week, \(p\) golfers in each group, and \(w\) weeks [1406.7196]. In the graph-theoretic and tournament-scheduling formulation, \(n\) denotes the number of players, \(k\) the group size, and one round is a partition of all \(n\) players into \(n/k\) groups of size \(k\), under the constraint that no pair of players meets more than once [2007.10704]. In the design-theoretic formulation, \(\text{SGP}(v,k,r)\) consists of arranging \(v=k\cdot n\) points into \(r\) rounds of \(n\) non-intersecting blocks of size \(k\) such that two points are assigned to the same block at most once [2507.23376].

These formulations are equivalent at the level of incidence structure. The set-constraint model writes group variables \(G_{i,j}\subseteq \mathcal P\), where \(i\in[1..w]\) is the week and \(j\in[1..g]\) is the group number, with the core conditions
\[
\forall i \in [1 .. w],\ \forall j \in [1 .. g],\ |G_{i,j}| = p \tag{9}
\]
and
\[
\forall i \in [1..w]\ \bigcup_{j=1 .. g} G_{i,j} = \mathcal{P}. \tag{10}
\]
The disjointness condition
\[
\forall i \in [1 .. w]\ \bigcap_{j=1 .. g} G_{i,j} = \emptyset \tag{11}
\]
is explicitly stated to be implied by (9) and (10), and therefore unnecessary in that model [1406.7196].

The graph-theoretic formulation replaces players by vertices of \(K_n\), groups of size \(k\) by copies of \(K_k\), and a round by a \(K_k\)-factor, that is, \(n/k\) vertex-disjoint copies of \(K_k\) covering all vertices [2007.10704]. After a round is played, the edges inside those \(K_k\)’s are deleted from the feasibility graph. Equivalently, if \(T=(H_1,\dots,H_r)\) is a tournament with \(H=K_k\), then the feasibility graph after rounds \(H_1,\dots,H_r\) is
\[
G = K_n \setminus \bigcup_{i\le r} H_i.
\]
A next round is feasible if and only if \(G\) still contains a \(K_k\)-factor [2007.10704].

A central quantitative restriction is the pair-count or point-degree bound. Since each point meets \(k-1\) others in each round and no pair may be repeated, the number of rounds satisfies
\[
r \le \left\lfloor \frac{v-1}{k-1}\right\rfloor,
\]
which is later denoted
\[
r'(v,k)=\left\lfloor\frac{v-1}{k-1}\right\rfloor.
\]
If a schedule attains this bound, it is maximal [2507.23376]. This upper bound also appears in the algorithmic literature as
\[
\mathrm{OPT}\le \frac{n-1}{k-1},
\]
for the maximum possible number of rounds in the classical offline SGP maximization problem [2007.10704].

## 2. Classical offline SGP and its design-theoretic structure

In the classical offline setting, all rounds are designed jointly. The design-theoretic interpretation is that SGP schedules are resolvable packings, and when every pair occurs exactly once they are resolvable block designs [2507.23376]. In particular, an RBIBD\((v,k)\) gives a schedule in which every pair of points occurs exactly once, and because the blocks resolve into parallel classes, each parallel class is a round. Such a design yields exactly
\[
\frac{v-1}{k-1}
\]
rounds, matching the upper bound, and is therefore maximal [2507.23376].

A resolvable group divisible design also yields SGP schedules, but with a different incidence pattern. A \(k\)-RGDD on \(v\) points with group size \(g\) gives
\[
\frac{v-g}{k-1}
\]
rounds, because each point meets all points outside its own group exactly once [2507.23376]. This explains an important distinction between two types of complete-looking schedules. In an RBIBD-derived schedule, every pair appears exactly once. In an RGDD-derived schedule, some pairs never occur because pairs within a design-group are never scheduled together [2507.23376].

The literature distinguishes carefully between *optimal* and *maximal*. A solution with \(r\) rounds is called optimal if no solution with more than \(r\) rounds is known, whereas it is maximal if no solution with more than \(r\) rounds is possible [2507.23376]. This distinction matters because many constructions for larger block sizes are presented as best-known schedules rather than as schedules certified to meet the absolute upper bound.

A common misconception is to identify any strong constructive schedule with a maximal one. The design literature separates these notions explicitly. For example, for \(k=4\), a URD fallback can yield an optimal schedule with \((v-5)/3\) rounds in exceptional cases, while maximality would require \((v-2)/3\) rounds [2507.23376]. This suggests that construction quality and extremal optimality should be distinguished even when the gap is only one round.

## 3. Constructive families and existence results

Resolvable combinatorial designs supply the strongest known exact constructions for many parameter ranges [2507.23376]. For block size \(k=3\), the paper states that maximal allocations exist for all \(v\equiv 3\pmod 6\) via Kirkman triple systems, and for all \(v\equiv 0\pmod 6\), \(v\ge 18\), via nearly Kirkman triple systems, which are \(3\)-RGDDs of type \(2^{v/2}\) [2507.23376]. The small exceptions are handled separately: for \(v=6\), one round of two blocks; for \(v=12\), RTD\((3,4)\) [2507.23376].

For block size \(k=4\), the construction theory is nearly complete. If \(v\equiv 4\pmod{12}\), an RBIBD\((v,4)\) gives \((v-1)/3\) rounds and is maximal. If \(v\equiv 0\pmod{12}\), a \(4\)-RGDD of type \(3^{v/3}\) gives \((v-3)/3\) rounds, equal to \(\lfloor (v-1)/3\rfloor\), and is therefore maximal. If \(v\equiv 8\pmod{12}\), a \(4\)-RGDD of type \(2^{v/2}\) gives \((v-2)/3\) rounds, also maximal when it exists. For the larger exceptional values, the cited work uses a URD\((\{2,4\};v)\) with
\[
r_2=4,\qquad r_4=\frac{v-5}{3},
\]
and the \(4\)-block rounds form an optimal schedule one short of maximal [2507.23376].

For \(k=5\), the situation is weaker. If \(v\equiv 5\pmod{20}\) and RBIBD\((v,5)\) exists, then one obtains a maximal allocation with \((v-1)/4\) rounds; existence is stated except for
\[
v\in\{45,345,465,645\}.
\]
Outside this RBIBD range, the paper often uses RTD\((5,v/5)\), giving \(v/5\) rounds, together with several special cases such as URD\((\{2,5\};30)\) with \(r_5=6\), an incomplete group divisible design for \(v=50\) with \(8\) rounds, and an analogous design for \(v=110\) with \(19\) rounds [2507.23376].

Two further constructive mechanisms broaden the available range. First, if there exist \(g\) MOLS\((n)\) with \(v=kn\), then the Sharma–Das construction gives a \((g+1)\)-round \((v,k)\) allocation, and if \(k\mid n\), the unused cliques contribute extra rounds, yielding
\[
\operatorname{optR}(n,k)+g+1.
\]
Second, the starter-block construction denoted \(\mathrm{ownSG}(v,k)\) develops modular starter blocks into parallel rounds, with unused pairs forming \(k\) cliques that can contribute extra rounds when \(k\mid (v/k)\) [2507.23376].

The 2025 classification states that a complete set of solutions is provided for up to 150 players. In that appendix-level sense, “complete set” means a full classified table of best-known constructions and round counts for all instances in the stated range, not that every one of those instances is maximal [2507.23376].

## 4. Greedy and online SGP

The 2020 work studies a different problem: not the classical offline existence question directly, but a greedy or online guarantee problem motivated by Swiss-system tournament scheduling [2007.10704]. Rounds are selected one by one, adaptively, with no preplanning and no ability to revise earlier rounds. The algorithm starts with \(G=K_n\), repeatedly chooses an arbitrary feasible \(K_k\)-factor, deletes its clique edges, and stops when no \(K_k\)-factor remains [2007.10704].

Its principal theorem is that for each \(n\in\mathbb N\) and \(H=K_k\), the greedy algorithm outputs a tournament with at least
\[
\left\lfloor \frac{n}{k(k-1)} \right\rfloor
\]
rounds [2007.10704]. In scheduling language: for any number \(n\) of golfers divisible by \(k\), if in each week one chooses any feasible partition into groups of size \(k\), then one is guaranteed at least
\[
R \ge \left\lfloor \frac{n}{k(k-1)} \right\rfloor
\]
weeks before getting stuck [2007.10704].

This theorem is explicitly not a statement about the optimum number of rounds in the classical SGP. It is a worst-case lower bound for arbitrary greedy round-by-round scheduling [2007.10704]. That distinction is essential. In structured offline instances, a complete schedule may have as many as
\[
\frac{n-1}{k-1}
\]
rounds, which is roughly \(k\) times larger than the greedy guarantee [2007.10704]. The result is therefore best understood as a robust online guarantee, not as an exact characterization of optimal schedules.

The bound is tight for the greedy model. The paper proves that there are infinitely many \(n\in\mathbb N\) for which there exists a tournament that cannot be extended after
\[
\left\lfloor \frac{n}{k(k-1)} \right\rfloor
\]
rounds [2007.10704]. The tightness construction takes
\[
n=j\,k(k-1)
\]
for infinitely many suitable integers \(j\), chooses a set \(A\) with
\[
|A|=\frac{n}{k}+1,
\]
and schedules rounds so that no edge between \(A\) and \(V\setminus A\) is ever used. After \(n/(k(k-1))\) rounds, all edges inside \(A\) are exhausted, and any further \(K_k\)-factor would require more outside vertices than exist because
\[
(k-1)\left(\frac{n}{k}+1\right) > |V\setminus A|. 
\]
This gives exact tightness for infinitely many instances [2007.10704].

The same paper also gives a conditional characterization. Assuming the Equitable \(\Delta\)-Coloring Conjecture, known for \(k\le 4\), a tournament with
\[
\left\lfloor \frac{n}{k(k-1)} \right\rfloor
\]
rounds is non-extendable if and only if \(K_{n/k+1}\) is a subgraph of the complement \(\bar G\) of the feasibility graph [2007.10704]. Under that conjecture and when \(n>k(k-1)\), choosing the last greedy round carefully can always yield one additional round:
\[
\left\lfloor \frac{n}{k(k-1)} \right\rfloor + 1.
\]
Because this improvement is conditional, the main unconditional statement remains the lower bound above [2007.10704].

## 5. Algorithmic formulations: SAT, set constraints, and approximation

The set-constraint approach models SGP declaratively, then automatically encodes the model into SAT [1406.7196]. The central idea is that a group is a set of golfers and a weekly schedule is a matrix of such set variables \(G_{i,j}\). This eliminates positional variables, because order inside a group is irrelevant [1406.7196]. The socialization condition is expressed either by an implication form,
\[
\forall w_1,w_2 \in [1..w],\ p_i,p_j \in \mathcal{P},\ g_1,g_2 \in [1..g],\ 
w_1 > w_2 \ \wedge\ i > j \ \wedge\ g_1 \ge g_2 \ \wedge\ p_i \in G_{w_1,g_1}
\]
\[
\wedge\ p_j \in G_{w_1,g_1} \ \wedge\ p_i \in G_{w_2,g_2} \ \rightarrow\ p_j \notin G_{w_2,g_2} \tag{12}
\]
or by the bounded-intersection form
\[
\forall w_1,w_2 \in [1..w],\ g_1,g_2 \in [1..g],\ 
w_1 > w_2 \ \wedge\ g_1 > g_2 \ \wedge\ |G_{w_1,g_1} \cap G_{w_2,g_2}| \le 1. \tag{13}
\]
The paper keeps the implication formulation for experiments because it yields lower asymptotic complexity [1406.7196].

The encoding framework is generic for set CSPs. It introduces one Boolean support variable for each element of the support of each set, and supplies generic encoding rules for membership, equality, inclusion, union, intersection, multi-union, multi-intersection, difference, and cardinality [1406.7196]. Cardinality is encoded using the unary totalizer/comparator encoding of Bailleux and Boufkhad, rather than a combinatorial subset encoding considered too large [1406.7196].

The reported asymptotic complexity of the implication-based socialization encoding is
\[
O(w^2.g^4.p^2),
\]
whereas the cardinality-intersection formulation has
\[
O(w^2.g^5.p^3).
\]
This is the stated reason for preferring (12) over (13) in the implementation and experiments [1406.7196].

Symmetry breaking is treated in two distinct ways: by adding constraints and by modifying the model itself [1406.7196]. The first strategy fixes the first week in a canonical way and then spreads the first \(p\) golfers across future weeks. As added constraints, these are
\[
\forall i \in [1 .. p.g],\ p_i \in G_{1,\ i \ div\ (p+1)} \tag{17}
\]
and
\[
\forall i \in [2 .. w],\ \forall j \in [1 .. p],\ p_j \in G_{i,j}. \tag{18}
\]
As model refinements, the same symmetries are handled by shrinking supports and redefining variables, thereby reducing SAT instance size before clause generation [1406.7196].

The comparative terminology used in that paper is: DE for direct encoding, TME for the Triska-Musliu improved hand-written encoding, TMESB for TME with symmetry breaking, SCE for set-constraint encoding, SCESBC for SCE with symmetry breaking by adding constraints, and SCESBM for SCE with symmetry breaking by modifying the model [1406.7196]. The paper claims and experimentally supports that the automatically generated SAT instances have fewer clauses than the improved hand-written encoding, and that after unit propagation they also use fewer variables while solving faster on hard instances [1406.7196].

The greedy scheduling paper gives a different algorithmic perspective. With the upper bound
\[
\mathrm{OPT}\le \frac{n-1}{k-1}
\]
and the lower bound
\[
R_{\text{greedy}} \ge \left\lfloor \frac{n}{k(k-1)} \right\rfloor,
\]
it derives
\[
R_{\text{greedy}} \ge \frac{1}{k}\,\mathrm{OPT} -1,
\]
and also states the explicit approximation ratio
\[
\frac{k-1}{2k^2-3k-1}.
\]
Accordingly, the greedy algorithm is interpreted as a polynomial-time constant-factor approximation algorithm for the SGP maximization problem [2007.10704].

## 6. Proof mechanisms, examples, and limitations

The proof of the greedy lower bound is based on a complement-coloring argument using the Hajnal–Szemerédi theorem [2007.10704]. After \(r\) rounds, every vertex has lost exactly \(r(k-1)\) incident edges, so
\[
\deg_G(v)=n-1-r(k-1).
\]
If
\[
r\le \left\lfloor \frac{n}{k(k-1)} \right\rfloor -1,
\]
then
\[
\deg_G(v)\ge n-1-\left(\left\lfloor \frac{n}{k(k-1)} \right\rfloor -1\right)(k-1) \ge n-1-\frac{n}{k}+k-1,
\]
hence in the complement graph
\[
\Delta(\bar G)\le \frac{n}{k}-k+1.
\]
With \(\ell=n/k\), this gives \(\Delta(\bar G)\le \ell-1\), so Hajnal–Szemerédi yields an equitable \(\ell\)-coloring of \(\bar G\), and since the color classes all have size exactly \(k\), this corresponds to a \(K_k\)-factor in \(G\). Therefore the algorithm cannot stop before round
\[
\left\lfloor \frac{n}{k(k-1)} \right\rfloor
\]
[2007.10704].

The design-theoretic literature provides explicit examples illustrating the contrast between RBIBD and RGDD schedules. A schedule with \(v=28\), \(k=4\), \(7\) blocks per round, and \(9\) rounds is maximal because
\[
\left\lfloor\frac{28-1}{4-1}\right\rfloor=9,
\]
and every pair appears exactly once, so it comes from an RBIBD\((28,4)\) [2507.23376]. By contrast, the \(v=24\), \(k=4\) example has \(6\) blocks per round and \(7\) rounds, is also maximal because
\[
\left\lfloor\frac{24-1}{3}\right\rfloor=7,
\]
but does not realize all pairs because it comes from an RGDD rather than an RBIBD [2507.23376].

The greedy theorem can also be instantiated numerically. For \(n=12\), \(k=3\), it guarantees
\[
R \ge \left\lfloor \frac{12}{3\cdot 2} \right\rfloor = 2,
\]
while the pair-count upper bound gives \(\mathrm{OPT}\le 5\) rounds integrally. For \(n=20\), \(k=4\), the unconditional worst-case guarantee is only
\[
\left\lfloor \frac{20}{4\cdot 3} \right\rfloor = 1,
\]
and for \(n=36\), \(k=4\), it is
\[
\left\lfloor \frac{36}{12}\right\rfloor = 3.
\]
These examples are presented specifically to show that the greedy theorem is a safe guaranteed minimum rather than a prediction of typical or optimal performance [2007.10704].

The SAT literature highlights a different practical limitation: complete symmetry breaking is possible in principle but usually impractical due to too many constraints [1406.7196]. It also notes that support size strongly affects SAT instance size, although for SGP the basic supports are already minimal [1406.7196]. The design-theoretic literature, in turn, is strongest for \(k=3\) and \(k=4\), weaker for larger \(k\), where many instances are handled by best-known constructions rather than complete maximality theorems [2507.23376].

Taken together, these strands show that SGP is not a single research problem but a family of related questions. One line studies maximal offline schedules through RBIBDs, RGDDs, URDs, MOLS, RTDs, and explicit constructions [2507.23376]. Another studies declarative modeling and automated SAT compilation for exact search on fixed parameter triples \(g\)-\(p\)-\(w\) [1406.7196]. A third studies worst-case guarantees for arbitrary greedy round-by-round scheduling and turns that guarantee into a constructive approximation algorithm [2007.10704]. This suggests that the term “Social Golfer Problem” now denotes both a classical extremal design problem and a broader algorithmic domain spanning exact modeling, constructive design theory, and online scheduling.

Source: https://www.emergentmind.com/topics/social-golfer-problem-sgp