A Resource Allocation Tournament

1 Introduction

• The allocation problem also arises beyond sports, including firms distributing budgets across projects, parties allocating campaign resources, and countries maintaining advantage in arms races.The introduction presents these alongside sports as settings where resources are distributed strategically.Source: §1 · pp. 1–2
• Simultaneous allocation means committing to allocations for all stages before observing earlier outcomes, unlike sequential allocation chosen stage by stage.Examples include advance planning of training effort and roster or budget allocations before a sports season begins.Source: §1 · p. 2
• Resource allocation contests constrain players to spend their entire budgets and maximize expected rewards, whereas standard profit-maximization contests subtract chosen costs from rewards.In contests, all participants bear their action costs even if they do not win.Source: §1 · p. 3
• The paper introduces a multiplayer round-robin resource allocation tournament with simultaneous all-pay matches, where players bear costs regardless of match outcomes.The framework models teams distributing costly resources across matches to compete for a prize.Source: §1 · pp. 1, 3, 4
• Prior round-robin tournament models surveyed in the introduction do not combine multiplayer competition with resource budgets as this model does.The paper states that its model has not previously been studied.Source: §1 · p. 4

2 Model

• In an n-player round robin, each player allocates one resource unit across n−1 matches simultaneously and in advance, without observing opponents, match order, or outcomes.The strategy is an allocation vector whose entries sum to one.Source: §2 · pp. 4–5
• A player wins a match by allocating more resources and earns 1 point; equal allocations produce a draw worth 0.5 point to each player.Source: §2 · p. 5
• The total-points scheme pays each player their accumulated points, while winner-takes-all shares one unit equally among players tied for the highest tournament total.Source: §2 · p. 5
• Expected payoffs average over all opponent and match-order permutations; the analysis seeks symmetric Nash equilibria and starts with three players because the two-player case is trivial.Expected total points are evaluated under the total-points scheme, and winning probability under winner-takes-all.Source: §2 · p. 5

• Every pure allocation strategy is a symmetric equilibrium in the three-player total-points game.The theorem states this for allocations of the form (x, 1 − x), with the domain for x omitted in the supplied passage.Source: §3 · pp. 5–6
• For any response strategy (y, 1 − y), the payoff is P0 = 1, independent of the response chosen, and therefore that strategy is also a best response.Three matches take place, so P0+2P1 = 3; in a symmetric equilibrium, P0 = P1 = 1.Source: §3 · p. 6
• Every pure-strategy profile is an equilibrium, and any mixed strategy constitutes a symmetric equilibrium.This extends the result beyond profiles in which all players use the same pure strategy.Source: §3 · p. 6

• In the four-player game, no pure symmetric equilibrium exists: shifting a small amount from the largest allocation to the other two improves payoff.Source: §4 · p. 7
• Better responses can cycle: against (0.34, 0.33, 0.33), a player can switch to (0.43, 0.43, 0.14), then (0.53, 0.24, 0.23), before returning to the initial allocation.Source: §4 · p. 7
• For an n+1-player game, any Blotto strategy is a mixed symmetric equilibrium, and profiles where players use different Blotto strategies are also equilibria.Source: §4 · pp. 8–9

5 Winner-takes-all payoff: three players

• The equal allocation (1/2, 1/2) is the unique symmetric equilibrium in the three-player winner-takes-all game.Against two equal-split opponents, deviating to an unequal split yields payoff 0, while equal allocation yields 1/3.Source: §5 · pp. 9–10
• For pure strategies ranked x ≥ y ≥ z ≥ 1/2, the payoff characterization gives player payoffs that sum to 1.The complete payoff function is presented in Table 1.Source: §5 · p. 9, Table 1 · p. 9
• Symmetric mixed strategies of the form (X, 1 − X) do not constitute equilibria, though their best responses can be characterized.The paper specifically notes that even a uniform distribution over X is not a symmetric equilibrium.Source: §5 · p. 10
• Against opponents independently drawing allocations from a continuous density, a deterministic split is a best response exactly when Pr(1 − y < X < y) = 0; equal allocation is always a best response.This condition applies to a deterministic allocation (y, 1 − y) with y ∈ [1/2, 1].Source: §5 · pp. 10–11
• Against two opponents drawing allocations uniformly, the unique best response is equal allocation (1/2, 1/2).The example considers deterministic choices (y, 1 − y) with y ∈ [1/2, 1].Source: §5 · p. 11

6.1 Geometric representation

• Ordering allocations as x ≥ y ≥ z reduces the four-player strategy space to a triangle with vertices (1, 0, 0), (1/2, 1/2, 0), and (1/3, 1/3, 1/3).The ordering removes strategically equivalent representations.Source: §6.1 · pp. 11–12

6.2 Numerical discretization

• With δ = 0.01, the four-player discretized strategy space contains 884 pure strategies.The number of grid strategies grows on the order of 1/δ^2.Source: §6.2 · p. 13

6.3 Pure strategies

• No pure strategy is a symmetric equilibrium in the continuous four-player game: against any common strategy, a slight reallocation yields a strictly higher payoff than 0.25.Against (x, y, z), the deviation (x − 2ε, y + ε, z + ε) is profitable; the result is independent of the numerical discretization.Source: §6.3 · pp. 13–14

6.4 Mixed strategies

• On the grid, three mixed-strategy examples have best-response payoffs of 0.362, 0.302, and 0.299; these are closer to equilibrium than any pure strategy, but not sufficiently close.The first example samples uniformly from a strategy triangle, and the second assigns equal probability to two line segments.Source: §6.4 · p. 15

6.5 Algorithmic solution

• The proposed approximate best-response dynamics with stochastic inertia combines inertia, approximate best responses, and stochastic approximation to synthesize their strengths, improve performance, and encourage convergence in the game.In this game, it seeks a 0.01-equilibrium, corresponding to a best-response payoff below 0.26 because the equilibrium payoff is 0.25. Candidate responses’ expected payoffs are approximated by simulating 10,000 realizations of the other players’ strategy combinations.Source: §6.4 · pp. 14–15, §6.5 · pp. 15–16
• With one million simulations, the returned four-player strategy has a best-response payoff of around 0.254, close to the equilibrium payoff of 0.25.In 1,000 iterations, the algorithm first found a strategy below 0.26 after approximately 200 iterations; its best strategy during the run had a payoff around 0.257.Source: §6.5 · pp. 16–17, Figure 3 · p. 17

6.6 Five players

• In the five-player game, the algorithm reached a strategy with payoff below 0.21 after approximately 1,250 iterations; the equilibrium payoff is 0.20.A strategy below 0.215 was first reached after approximately 250 iterations.Source: §6.6 · p. 19

7 Conclusion

• Across four variations, symmetric equilibria are pure for three players and mixed for four or more; the authors also proposed best-response dynamics and used them to solve one variation.The symmetric equilibrium strategies are summarized in Table 3.Source: §7 · p. 19
• Total points represents a prize for each match win, whereas winner-takes-all awards a prize only for winning the tournament.Source: §7 · p. 19
• With four or more players under total points, every Blotto strategy is a symmetric equilibrium, but whether non-Blotto symmetric equilibria exist remains open.The equilibrium is not unique because many different Blotto strategies are equilibria.Source: §7 · p. 20
• Under winner-takes-all with three players, the authors prove a unique symmetric equilibrium that is a best response against any common strategy of the other two players, but may not be a best response in the general case.For the stated x > y > 0.5 case, the passage says it is more beneficial for the third player to choose z = y than z = 0.5.Source: §7 · pp. 20–21
• After 5,000 iterations, decreasing-tolerance dynamics returned a best-response payoff of 0.252, cutting 50% of the equilibrium-payoff gap, versus 0.254 after 1,000 iterations with fixed γ = 0.01.The authors note that further parameter optimization may improve the result.Source: §7 · p. 21
• Future work could allow stage-by-stage adaptive allocations, combine tournament-win rewards with a last-place penalty or use Tullock match outcomes, and study asymmetric budgets with combined tournament and match-win prizes.The paper's model assumes players allocate resources before competition begins.Source: §7 · p. 21

A Proof of Lemma 8

• The payoff table covers all 32 case–configuration combinations except the trivial case x = y = z.The table is continued in the following passage.Source: A Proof of Lemma 8 · pp. 21–22
• Player 1 allocates x against Player 2 in configurations A, B, C, and D.Source: A Proof of Lemma 8 · p. 21
• Player 2 allocates y against Player 1 in configurations A, C, E, and G.Source: A Proof of Lemma 8 · p. 21
• Player 3 allocates z against Player 1 in configurations A, B, E, and F.Source: A Proof of Lemma 8 · p. 21

B Proof of Remark 10

• Each opponent’s allocation faced by the player follows g_X = (f_X + f_{1−X})/2, which is symmetric around 1/2.The probability of facing X or 1−X is 1/2 each, and the two opponents’ realizations are independent.Source: B Proof of Remark 10 · p. 23
• The player wins outright with probability a_y(1−a_y); a triple draw occurs with probability 1/2 and pays 1/3.With probability 1, all allocations are distinct, so matches have no draws; a tournament triple draw occurs when each player wins one match.Source: B Proof of Remark 10 · p. 23
• Triple draws arise from two cyclic match-outcome patterns; the first has probability a_y^2, while the second is split into three disjoint allocation cases.The proof sums the cases’ probabilities to obtain a total triple-draw probability of 1/2.Source: B Proof of Remark 10 · p. 23
• The payoff is P0 = 1/6 + ay(1−ay), where ay = Pr(X > y | X ∼ gX).Source: B Proof of Remark 10 · p. 23

C Discretized version of the returned strategy

• Figure 7 shows the normalized and discretized probability weights assigned by the returned strategy.Source: Figure 7 · p. 24