Abstract
In the philosophy of science, the problem of underdetermination poses a significant challenge: multiple theories can often explain the same set of observations despite having different internal structures. While scientists intuitively prefer "coherent" theories over ad-hoc ones, coherence has remained a largely subjective concept. This paper proposes a formal, logic-based framework to quantify the coherence of scientific theories. By defining the "utility" of individual laws based on their frequency of use in deriving observations, the authors provide a mathematical metric to rank empirically equivalent theories. The framework is validated through case studies, including the "Black Swan" problem and Craig’s Trick, demonstrating that coherence naturally penalizes patchwork explanations and rewards structural unity.
Introduction
When two scientific theories make identical empirical predictions, how do we determine which is superior? This phenomenon, known as underdetermination, suggests that empirical evidence alone is sometimes insufficient for theory selection. Historically, researchers have relied on qualitative virtues such as simplicity, elegance, and coherence. However, without a rigorous definition, "coherence" remains a matter of aesthetic preference rather than scientific measurement.
This paper addresses this gap by introducing a quantitative measure of coherence. The authors argue that a theory is coherent when its constituent laws work together harmoniously to explain data, rather than acting as a collection of disjointed, independent rules. By formalizing this intuition, the paper provides a tool to distinguish between theories that are empirically identical but structurally distinct.
Methodology: Quantifying Coherence
The authors' framework is built upon a logic-based approach that evaluates how individual formulas within a theory contribute to its overall explanatory power. The process involves three core components:
- Support Sets: For a given set of inputs ($I$) and observed outputs ($O$), a support set is defined as the smallest subset of a theory ($T$) that can logically derive the observations. This ensures that only the necessary components of a theory are considered for any specific explanation.
- Utility of a Formula: The utility of a specific law or formula within a theory is its relative frequency of appearance across all possible support sets for a given observation. A law that is essential for explaining a wide variety of phenomena has high utility, whereas a law used only for a single, specific case has low utility.
- The Coherence Metric: The coherence of the entire theory is calculated as the average utility of all its constituent formulas.
This methodology improves upon previous attempts at formalizing coherence by "relativizing" the support to specific input sets. This mirrors the actual practice of science, where theories are tested under specific experimental conditions (e.g., "Given these initial conditions, does the theory predict this specific result?").
Evaluation and Case Studies
As a theoretical work in logic and philosophy, the paper evaluates its metric using formal "toy models" and historical scientific scenarios rather than empirical datasets. These benchmarks demonstrate the metric's ability to identify structural flaws in theories.
The Black Swan Fix
This case study compares two ways of handling anomalous data. Imagine a theory stating "all swans are white" is confronted with $n$ observations of black swans.
- The Ad-Hoc Approach: One could simply add $n$ specific rules to the theory, each accounting for one specific black swan. The authors show that in this scenario, the coherence of the theory trends toward zero as more data arrives, represented by the formula $\frac{2k}{2n+1}$.
- The Theoretical Approach: Alternatively, one could introduce a new theoretical category (e.g., "Australian swans"). This maintains a stable, significantly higher coherence (0.5 in the authors' example), proving that introducing unified theoretical terms is structurally superior to listing exceptions.
Craig’s Trick
In the philosophy of science, "Craig’s Trick" is a logical maneuver used to argue that theoretical (unobservable) terms are unnecessary because they can be stripped away while maintaining empirical accuracy. The authors apply their metric to this maneuver and find that the resulting "stripped" theory is highly incoherent. This provides a formal, mathematical justification for why scientists reject such models in favor of those that include unobservable theoretical entities.
Mendel’s Laws
The paper cites Mendel’s Laws of inheritance as a benchmark for high coherence. In Mendelian genetics, a small set of independent theoretical characteristics accounts for a vast and diverse array of biological observations, resulting in high utility for each law and, consequently, a high coherence score for the theory as a whole.
Key Insights
- Coherence as Functional Integration: A theory is coherent if its laws are "tightly coupled." If Law A only explains Observation 1 and Law B only explains Observation 2, the theory is a mere collection of facts. High coherence requires that laws work in concert across multiple observations.
- The Ad-Hoc Penalty: The mathematical framework naturally penalizes "patchwork" theories. By requiring laws to have high utility across support sets, the metric identifies when a theory is being artificially inflated with specific rules to cover gaps in its explanatory power.
- Modularization: The metric suggests that scientific knowledge is often divided into modules because specific groups of laws support specific groups of observations. Coherence can thus explain the natural boundaries between different scientific disciplines or sub-fields.
Limitations and Future Directions
While the proposed metric offers a robust starting point, the authors acknowledge several limitations:
- Clausal Basis: The model currently assumes theories are represented as "clauses" in formal logic, which may not fully capture the linguistic complexity of all scientific fields.
- Computational Complexity: Calculating all possible support sets for a complex, real-world theory is computationally intensive, potentially limiting the metric's immediate application to large-scale theories.
Future research will explore the relationship between coherence and Kolmogorov complexity—the measure of the computational resources needed to specify an object. The authors hypothesize that maximally coherent theories may also be the most compressed representations of data. Additionally, they aim to investigate how the drive for coherence influences the historical evolution and modularization of scientific knowledge.
Summary updated