Subgame Perfect Equilibrium in Extensive Games

• Subgame Perfect Equilibrium (SPE) is a refinement of Nash equilibrium that requires every subgame strategy profile to be optimal, ensuring sequential rationality and credibility.
• Backward induction and transfinite induction are used to construct SPE, especially in infinite games where quasi-profiles partition the game tree for equilibrium extension.
• Structural conditions like the 'disjoint-chain' prohibition ensure the existence of global-Pareto optimal SPE, highlighting key links between preference orderings and equilibrium outcomes.

A subgame perfect equilibrium (SPE) is a refinement of Nash equilibrium applicable to extensive-form games, whereby a strategy profile constitutes a Nash equilibrium in every subgame. This concept ensures sequential rationality, ruling out non-credible threats by requiring that the prescribed strategies remain optimal even after any possible history. SPE is particularly significant in infinite games and games defined on graphs, where deviations at arbitrary histories must be considered, and where classical Nash equilibrium may allow for suboptimal or unstable behaviors off the equilibrium path.

1. Formal Definitions and Essential Properties

Consider a perfect-information extensive-form game defined as a tuple $g = \langle A,\;T,\;d,\;O,\;v,\;(\prec_a)_a\rangle$, where:

• $A$ is the player set.
• $T \subseteq \Sigma^*$ is the pruned tree of finite histories, with $succ(T,\gamma)$ the successors of history $\gamma$.
• $d: T \to A$ assigns to each $\gamma$ the moving player.
• $O$ is the set of outcomes.
• $v: [T] \to O$ is the (possibly measurable) outcome function over infinite plays.
• $\prec_a$ is a strict preference for player $A$0 on $A$1.

A strategy profile $A$2 selects at every node $A$3 a successor $A$4. For each subgame $A$5 (the subtree rooted at $A$6), the restriction $A$7 defines a strategy profile for $A$8.

A Nash equilibrium (NE) requires each player to have no profitable deviation at her own nodes in the root game. An SPE requires that $A$9 is a NE in every $T \subseteq \Sigma^*$0:

$T \subseteq \Sigma^*$1

This eliminates non-credible threats: every deviation, even far off the initial path, must be unprofitable.

2. Existence of SPE and Preference Structures

Existence results for SPE depend critically on the payoff structure and preference relations. In finite extensive-form games with well-founded, acyclic preferences, backward induction guarantees the existence of an SPE (0705.3316). This was generalized by Le Roux to infinite games with outcome functions measurable in the Hausdorff difference hierarchy (i.e., $T \subseteq \Sigma^*$2-measurable in the Baire space) (Roux, 2015).

For countable trees and finite $T \subseteq \Sigma^*$3, the main equivalence is:

• If for all $T \subseteq \Sigma^*$4, $T \subseteq \Sigma^*$5, the "disjoint chain" (SPE-killer) pattern is forbidden—namely, it is never the case that $T \subseteq \Sigma^*$6 and $T \subseteq \Sigma^*$7—then every game with linear preferences admits an (even global Pareto-optimal) SPE.

If preferences admit infinite ascending chains, or the disjoint chain condition is violated, existence of SPE may fail—even in games with mild discontinuities in payoffs or only a small violation in measurability.

3. Construction and Proof Techniques: Induction and Hierarchy

The proof of existence for infinite games with $T \subseteq \Sigma^*$8-measurable outcomes (Roux, 2015) relies on a transfinite induction over the Hausdorff difference hierarchy's complexity. At each complexity level, the construction uses "quasi-profiles"—multi-valued maps assigning to each node a subset of allowed successors (moves)—to partition the original game tree into smaller subgames. By constructing (Pareto-optimal) SPEs inductively on these lower-complexity subgames and patching them together, the equilibrium is extended to the larger game.

For successor and limit ordinals in the hierarchy, the proof exploits two-player antagonistic and quasi-antagonistic constructions on each partition class, recursively reinstating more preferred classes by backward induction.

4. The Disjoint-Chain (SPE-Killer) and Its Role

The central obstruction to universal existence of SPE in general preference structures is the presence of disjoint chains as follows. For linear preferences $T \subseteq \Sigma^*$9:

$succ(T,\gamma)$0

must hold for all players and all triples of outcomes. If this condition is violated, there exist games—regardless of the regularity of the payoff function—without any SPE, due to essentially incompatible preference cycles that cannot be resolved at any subgame, a phenomenon intimately connected to the failure of backward induction in certain partially ordered preference systems.

This marks a precise structural boundary for the general existence of SPE in games with linear preferences and measurable outcome functions (Roux, 2015).

5. Pareto-Optimality and Equilibrium Refinements

The construction in (Roux, 2015) achieves not merely existence of an SPE but of a global-Pareto SPE. That is, at each node of the game, the equilibrium outcome is Pareto-optimal among all outcomes occurring in the corresponding subgame—no player can be made better off without making another worse off among reachable outcomes in the subtree.

The method for refining outcomes to Pareto-optimality involves, at each stage of the hierarchy, selecting as equilibrium from available options at each node those that maximize the set of outcomes not strictly dominated for some agent, using the quasi-profile partitioning structure.

6. Beyond Linear Preferences: Extensions and Limitations

For preferences that are not linear, additional complications arise. While the constructive scheme can be extended to certain non-linear (partial order) preferences, existence of SPE may fail without the totality and proper hierarchical structure required to preclude the existence of infinite improvement cycles or the SPE-killer configuration.

With abstract atomic outcomes and arbitrary preference relations, as formalized in (0705.3316), the equivalence of acyclicity of preferences, existence of NE, and existence of SPE can be fully characterized and is constructively formalized in proof assistants such as Coq. Specifically, the equivalence between the absence of cyclic preferences and the ability to use backward induction to construct SPE holds in an entirely general framework—highlighting the fundamental connection between order-theoretic properties of preferences and the algorithmic (backward induction) construction of equilibrium.

7. Implications and Applications

The identified conditions and constructive frameworks for SPE have direct application in extensive-form infinite games, particularly in economic, algorithmic, and verification contexts where payoffs may be discontinuous or preference structures complex. The results delineate precisely when backward induction and its infinite analogues are valid, providing a robust theoretical underpinning for equilibrium selection in multi-player settings with non-wellfounded or non-continuous outcomes. The "disjoint-chain" condition serves as a structural descriptor for the design of preference systems if universal existence of global-Pareto subgame perfect equilibria is desired (Roux, 2015, 0705.3316).