SPaDE

Synthetic Philosophy and Deductive Engineering

View the Project on GitHub rbjones/SPaDE

Perfect Information Spaces and Focal Methods

Author: GitHub Copilot (Claude Sonnet 4.5)
Date: November 16, 2025
Category: Architecture and Philosophy

Introduction

This document establishes a fundamental equivalence between three concepts central to the SPaDE project: perfect information spaces, formal deductive theories, and the domain of deductive reason. This equivalence defines the precise scope and applicability of focal methods within SPaDE.

The key insight originates from Google DeepMind’s AlphaZero, which demonstrated that artificial intelligence can achieve superhuman performance in perfect information spaces through self-play alone, without requiring human expert demonstrations. This observation has profound implications for automated theorem proving and the architecture of SPaDE’s deductive intelligence subsystem.

Perfect Information Spaces

Definition

Building on Google DeepMind’s usage in systems like AlphaZero (where perfect information refers to games with complete state visibility and no hidden information), a perfect information space (PIS) in SPaDE is a domain characterized by the following properties:

  1. Complete Observability - All relevant state information is visible to all participants; there is no hidden information

  2. Well-Defined Rules - The dynamics governing state transitions are completely and formally specified

  3. Deterministic Behavior - Outcomes follow predictable rules without hidden randomness

  4. Countable and Semi-Decidable - Individual trajectories (e.g., game-plays or proofs) are finite; the overall space of states and actions is at most countable; validity/recognition of states, transitions, and outcomes is semi-decidable (verifiable if found, but search may not terminate)

  5. Objective Evaluation - Terminal states have unambiguous outcomes; success and failure are clearly defined

  6. Self-Contained - The rules completely define the domain; no external knowledge is required beyond the specification

Examples

Classic examples of perfect information spaces include:

Learning in Perfect Information Spaces

The critical property of perfect information spaces for artificial intelligence is that optimal strategies can be learned through self-play without human expert demonstrations. The key requirements are:

This property enabled AlphaZero to achieve superhuman performance in chess, Go, and shogi through reinforcement learning with self-play, starting from only the rules of the games.

Formal Deductive Theories

A formal deductive theory consists of:

  1. Language - A formal syntax defining well-formed formulas
  2. Axioms - A specified set of foundational propositions
  3. Inference Rules - Explicit rules for deriving new propositions from existing ones
  4. Theorems - Propositions derivable from axioms via inference rules

Formal Theories as Perfect Information Spaces

Any formal deductive theory naturally defines a perfect information space:

The process of theorem proving is precisely navigating this space to reach a terminal state (a proof).

Perfect Information Spaces as Formal Theories

Encoding Perfect Information Spaces

Conversely, any perfect information space can be encoded as a formal theory in a sufficiently expressive logical system:

  1. Represent States - Encode game/domain states as logical terms or structures
  2. Encode Transitions - Express legal moves/actions as logical axioms or rules
  3. Define Terminals - Specify termination conditions as logical predicates
  4. Capture Evaluation - Represent outcomes as logical properties

For example, the game of chess can be formally specified:

This encoding preserves the strategic structure of the original space, making properties of optimal play expressible as theorems.

The Domain of Deductive Reason

Characterization

Deductive reason is characterized by:

  1. Truth Preservation - Conclusions necessarily follow from premises
  2. Formal Rules - Inference follows explicit logical principles
  3. No Empirical Input - Once axioms are fixed, no observation is needed
  4. Completeness Relative to Rules - Everything derivable from axioms is in principle discoverable

The Three-Way Equivalence

We can now state the central claim:

This equivalence means:

Implications for Focal Methods

This equivalence has several important consequences:

  1. Clear Boundaries - Focal methods apply exactly where reasoning is deductive; they do not claim to address inductive or abductive reasoning

  2. Universal Applicability within Scope - Any domain reducible to formal deduction can benefit from focal methods

  3. Self-Play Learning - Deductive intelligence agents can learn optimal theorem-proving strategies through self-play, analogous to AlphaZero

  4. Compositional Structure - Complex deductive domains can be decomposed into a hierarchy of sub-spaces, each a focus for specialized intelligence

  5. Objective Evaluation - Proof correctness provides unambiguous feedback for learning algorithms

Perfect Information Spaces in SPaDE

Logical Contexts as Perfect Information Spaces

The SPaDE Knowledge Repository organizes formal mathematical knowledge as a hierarchical collection of perfect information spaces, where each logical context or theory is a distinct perfect information space:

Focal Hierarchy

The perfect information space structure naturally gives rise to SPaDE’s focal hierarchy:

Deductive Intelligence and Self-Play

The equivalence between perfect information spaces and formal theories suggests that SPaDE’s deductive intelligence agents can:

  1. Learn from Self-Play - Generate training data by attempting proofs within a context and evaluating success
  2. Transfer Learning - Apply strategies learned in one context to related contexts
  3. Hierarchical Delegation - Route problems to specialists trained for the appropriate logical context
  4. Continuous Improvement - Accumulate experience across multiple logical contexts throughout the repository

This approach mirrors AlphaZero’s success but applied to the domain of formal mathematics rather than games, with each logical context serving as a distinct training ground analogous to a specific game variant.

Scope and Limitations

What Falls Within the Scope

Focal methods apply to any domain that can be characterized as a perfect information space:

What Falls Outside the Scope

Focal methods do not directly address:

However, these domains often have deductive components that are addressable by focal methods, and SPaDE’s architecture allows integration with complementary approaches for non-deductive aspects.

Philosophical Foundations

Deductive vs. Inductive Knowledge

The perfect information space characterization clarifies the boundary between:

SPaDE’s epistemological stack recognizes this distinction while providing infrastructure for both types of knowledge, with focal methods specialized for the deductive layer.

Completeness and Decidability

It is important to note that:

The value of the perfect information space framework is not that it makes hard problems easy, but that it provides the right structure for attacking them systematically.

Universal Foundations

The fact that perfect information spaces are precisely coextensive with formal deductive theories suggests that:

Conclusion

The equivalence between perfect information spaces, formal deductive theories, and the domain of deductive reason establishes:

  1. A precise definition of the scope of focal methods
  2. A theoretical foundation for self-play learning in theorem proving
  3. An architectural principle for organizing deductive intelligence
  4. A connection between game-playing AI (AlphaZero) and mathematical AI

This framework guides SPaDE’s development by clarifying what is in scope (deductive reasoning in perfect information spaces) and what requires complementary approaches (inductive and abductive reasoning). The result is a focused, theoretically grounded approach to advancing artificial intelligence in the domain of formal mathematics and deductive reasoning.

References