Math Optimization Game

Math Optimization & Game Theory

Mathematical optimization and strategic interaction theory, applied to reinforcement learning, governance convergence, and multi-agent coordination.

Routing

• USE FOR: Mathematical optimization, game theory, decision theory, mechanism design, multi-objective optimization, convex optimization, and reinforcement learning theory. Use for reasoning about Bellman equations, Nash equilibria, Pareto optimality, gradient methods, linear/convex/integer programming, and optimization-related mathematics. Covers both pure theory and computational methodology.
• DO NOT USE FOR: Lean optimization proofs (use @lean-math-optimization); strategy methodology (use @applied-strategy-analysis); general nonlinear dynamics (use @math-nonlinear-dynamics).
• TRIGGERS: optimization, game theory, decision theory, mechanism design, convex optimization, reinforcement learning.

Workflow

1. Classify the problem: convex / nonconvex / combinatorial / game-theoretic / mechanism-design / multi-objective / RL.
2. Pick the matching section of the body; identify Mathlib Convex / Optimization namespaces where applicable.
3. Produce the answer (algorithm, equilibrium, mechanism) with assumptions + complexity.
4. Hand off: to @lean-math-optimization for the Lean proof, to @math-nonlinear-dynamics if dynamics dominate, to @lean-zettelkasten.

Recovery & STOP

• STOP if the question is strategy-analysis (not optimisation) — delegate to @applied-strategy-analysis.
• STOP if a Mathlib pin-verified lemma is required — escalate to @lean-research.
• STOP if the problem requires empirical RL experiments — escalate to @research-council.

Handoffs

• Predecessors: agent:gateway, skill:lean-research.
• Successors: skill:lean-math-optimization, skill:lean-research, skill:lean-zettelkasten.

Part 1 — Optimization Theory Foundations

1.1 Convex Optimization

1.2 Linear & Integer Programming

• Simplex method: Pivoting through basic feasible solutions
• Interior point: Barrier methods for LP
• Branch & bound: Integer programming via LP relaxation
• Network flows: Min-cost flow, max-flow/min-cut
• Common application: pipeline composition as flow optimization

1.3 Nonlinear Optimization

Part 2 — Game Theory

2.1 Solution Concepts

2.2 Game Classes

• Zero-sum: $\sum_i u_i = 0$ — adversarial safety scenarios
• Potential games: $\exists \Phi$ s.t. $\Delta u_i = \Delta \Phi$ — alignment incentives
• Stochastic games: State evolves with actions — MDP foundation
• Mechanism design: Design rules to achieve desired outcomes — governance design
• Cooperative games: Coalition formation, Shapley value — trust allocation

2.3 Nash Existence & Computation

Existence (Nash 1950): Every finite game has at least one mixed-strategy NE.

Computation complexity: Finding NE is PPAD-complete (Daskalakis et al. 2009).

Algorithms:

• Support enumeration (small games)
• Lemke-Howson (bimatrix)
• Homotopy methods
• Learning dynamics (fictitious play, multiplicative weights)

Common formalization: trust dynamics as learning dynamics converging to NE-like fixed points.

Part 3 — Decision Theory

3.1 Expected Utility Theory

• Von Neumann-Morgenstern axioms → utility function
• Risk aversion: concave utility ($u'' < 0$)
• Quality gate thresholds as risk-adjusted decisions

3.2 Multi-Criteria Decision Making

3.3 Decision Under Uncertainty

• Maximin: Worst-case optimization — safety-critical governance
• Minimax regret: Minimize maximum regret — robust decisions
• Bayesian: Maximize expected utility with priors — adaptive trust
• Info-gap: Robustness to severe uncertainty — unknown-unknown resilience

Part 4 — Reinforcement Learning Theory

4.1 MDP Foundations

4.2 Key Theorems

• Bellman optimality: $V^(s) = \max_a [R(s,a) + \gamma \sum_{s'} P(s'|s,a) V^(s')]$
• Contraction mapping: $|TV - TV'|\infty \le \gamma |V - V'|\infty$
• Policy improvement: $\pi_{k+1}$ greedy w.r.t. $V^{\pi_k}$ ⟹ $V^{\pi_{k+1}} \ge V^{\pi_k}$
• Regret bounds: Policy gradient regret $\tilde{O}(\sqrt{T})$

4.3 Safe RL

• Constrained MDP: $\max_\pi V^\pi$ s.t. $C^\pi \le d$
• Lyapunov-based: Safety via Lyapunov barrier functions
• Shielding: Pre/post-safety filters on actions
• Common use: trust dynamics as safe RL with Lyapunov certification

Part 5 — Multi-Objective Optimization

5.1 Pareto Theory

• Pareto dominance: $x \prec y$ iff $f_i(x) \le f_i(y)$ for all $i$, strict for some
• Pareto front: Set of non-dominated solutions
• Scalarization: $\min_x \sum_i w_i f_i(x)$ recovers Pareto points (for convex)
• $\epsilon$-constraint: $\min f_1$ s.t. $f_j \le \epsilon_j$ for $j \ge 2$

5.2 Multi-Objective Formulation

Governance and safety problems are often inherently multi-objective:

• Maximize quality (CCV scores)
• Minimize risk (safety envelope violations)
• Maximize learning rate (nested learning convergence)
• Minimize computational cost (pipeline efficiency)

Part 6 — Host-Repository Lean Extension Points

Do not assume repository-local Lean modules, tactics, or namespaces exist unless the host repository explicitly provides them. Map local modules by role:

Part 7 — Research Methodology for Optimization

7.1 When to Apply This Skill

• Designing governance objective functions
• Analyzing convergence rates of iterative schemes
• Proving optimality of policy choices
• Designing multi-agent coordination mechanisms
• Analyzing trade-offs in safety vs performance

7.2 Key References

• Boyd & Vandenberghe (2004) — Convex optimization Bible
• Osborne & Rubinstein (1994) — Game theory foundations
• Puterman (2014) — MDP theory and algorithms
• Szepesvári (2010) — RL theory algorithms
• Miettinen (1999) — Nonlinear multi-objective optimization

7.3 Epistemic Mapping Targets