The Coherence Principle: A Falsifiable Prior for Model Selection from the Grammar of Theories

This paper proposes the "Coherence Principle," a falsifiable Bayesian framework for model selection that assigns prior probabilities based on a model's adherence to the validated structural "grammar" of existing theories, thereby penalizing unmotivated theoretical violations while remaining transparent and testable across historical and contemporary cases in physics.

The Gist

Imagine you are a judge in a courtroom, but instead of judging crimes, you are judging scientific theories. You have some new evidence (data) from the universe, and two different stories (models) are trying to explain it.

Usually, judges look at the evidence to decide who wins. But sometimes, the evidence is blurry, weak, or doesn't clearly point to one story over the other. In those moments, the judge has to rely on a "gut feeling" about which story seems more reasonable before even looking at the evidence. In science, this gut feeling is called a "prior."

The problem, according to this paper, is that scientists often use different, unspoken gut feelings. One scientist might think, "This theory is too complicated, so it's probably wrong" (Occam's Razor). Another might think, "This theory feels 'natural' because the numbers aren't weird." But these feelings are hard to measure and easy to argue about.

The Paper's Big Idea: The "Coherence Principle"
The authors propose a new, rule-based way to set that "gut feeling." They call it the Coherence Principle.

Think of established science (like the Standard Model of particle physics or General Relativity) as a giant, well-tested rulebook or a grammar for how the universe works. This rulebook contains things like:

• Energy must be conserved.
• Nothing can travel faster than light.
• The laws of physics should look the same everywhere (symmetry).

The Coherence Principle says: "If a new theory breaks these rules without a very good reason, it should start with a disadvantage."

How It Works: The "Grammar" Analogy

Imagine the laws of physics are like the rules of English grammar.

• The Background Theory (The Grammar): We know that sentences need a subject and a verb. We know you can't just swap nouns for verbs randomly. These are the "validated rules."
• The New Model (The Sentence): A scientist proposes a new theory.
  • Scenario A: The scientist writes a sentence that follows all the grammar rules perfectly. (e.g., "The cat sat on the mat.")
  • Scenario B: The scientist writes a sentence that breaks the rules for no reason. (e.g., "The cat jumped the blue yesterday.")

The Coherence Principle says: Scenario B gets a "penalty." It starts with a lower score because it violates the established grammar of the universe.

The "Coherence Cost"

The paper introduces a simple math trick to measure this penalty:

1. Count the Violations: Every time a new theory breaks a major rule (like breaking the speed of light or ignoring energy conservation) without a very strong, independent reason, it gets a "Coherence Cost" of 1.
2. The Penalty: The more violations, the lower the theory's starting score. The math looks like this: Score = e^(-Cost).
   • If you break 0 rules, your score is 1 (100% chance).
   • If you break 1 rule, your score drops to about 0.37 (37% chance).
   • If you break 2 rules, it drops to about 0.14 (14% chance).

This isn't about saying the theory is wrong; it's just saying, "Because you broke the rules, you need much stronger evidence to prove you're right."

Real-World Examples from the Paper

The authors test this idea on historical and current scientific debates:

1. The Neutrino Mystery (Current Day)

• The Situation: We know neutrinos have mass, but we don't know the exact pattern. One theory says all three types of neutrinos get their mass from the same "recipe" (Unified). Another says they get mass from three totally different, unrelated recipes (Disjoint).
• The Grammar Rule: The Standard Model likes "universality" (treating similar things the same way).
• The Result: The "Unified" theory follows the grammar. The "Disjoint" theory breaks the rule of universality. The Coherence Principle gives the Unified theory a slight head start. It doesn't solve the mystery alone, but it helps tip the scales when the data is weak.

2. Einstein vs. Newton (History)

• The Situation: In 1905, Einstein proposed General Relativity.
• The Grammar Check: If you looked at the rules of physics before 1905 (Newton's rules), Einstein's theory looked like a disaster. It broke the rules of "absolute time" and "instant gravity." The Coherence Principle would have said, "Einstein is breaking the grammar! Don't trust him!"
• The Twist: But by 1915, the "grammar" had changed. Special Relativity had already proven that "absolute time" was wrong. Once the grammar was updated to include "Lorentz invariance" (the new rule), Einstein's theory suddenly looked perfect (0 violations), while Newton's theory looked like it was breaking the new rules.
• Lesson: The principle works, but you must use the current rulebook, not an old one.

3. The "Ghost" in the Machine (Dark Energy)

• The Situation: Scientists are trying to figure out if the universe's expansion is speeding up in a weird way. Some theories suggest "phantom energy" that breaks the laws of stability (creating "ghosts" that make the universe unstable).
• The Result: The Coherence Principle says, "If your theory creates a ghost, you pay a penalty." It doesn't say the theory is impossible, but it says, "You need really, really strong data to prove this ghost exists."

What This Principle is NOT

The authors are very careful to say what this is not:

• It's not "Beauty": You can't just say, "This theory is beautiful, so it gets a bonus." The rules must be based on things we have actually tested and proven.
• It's not "Occam's Razor": Occam's Razor says, "Pick the simpler theory." The Coherence Principle says, "Pick the theory that fits the rules of the universe, even if it's complicated."
• It's not a magic wand: If the data is super clear and loud, the data wins. This principle only helps when the data is quiet and confusing.

The Bottom Line

The Coherence Principle is a way to make scientists' "gut feelings" about which theories are plausible transparent and fair.

Instead of saying, "I feel like this theory is weird," a scientist can now say, "This theory breaks three specific, well-tested rules of physics, so I am giving it a penalty of 0.05."

It turns the invisible bias of "trust in established science" into a visible, calculable number. It reminds us that while science is always looking for new things, it should be harder to believe in things that break the rules we've already proven to be true.

Technical Summary

Technical Summary: The Coherence Principle

1. Problem Statement

In modern cosmology and particle physics, Bayesian model selection often faces a critical methodological challenge: when empirical data are weak, indirect, or close to physical boundaries, the posterior odds between competing hypotheses inherit a strong, frequently unacknowledged dependence on the prior assigned to the model space. Standard approaches to defining these priors—such as reference priors, hierarchical priors, or appeals to naturalness—either ignore relevant theoretical knowledge or rely on criteria that are difficult to define operationally.

Working physicists routinely assign prior plausibility based on a model's structural compatibility with the "validated architecture" of successful background theories (e.g., symmetries, conservation laws, locality). However, this judgment is rarely formalized as a quantitative, reproducible component of Bayesian inference. The paper identifies a gap between this common physical reasoning and the rigorous requirements of Bayesian model selection, particularly in regimes where the Bayes factor alone is insufficient to distinguish between models (e.g., neutrino mass ordering, dark energy equation of state).

2. Methodology: The Coherence Principle

The authors propose the Coherence Principle, a general procedure for assigning model priors based on the "structural grammar" of a successful background theory.

2.1 Structural Grammar

The method begins by defining a background theory $T$ and its grammar $G(T)$, a set of validated structural rules $\{p_1, p_2, \dots, p_n\}$. These rules are coarse-grained architectural features such as:

• Symmetries: Gauge invariance, Lorentz invariance, family universality.
• Spacetime Principles: Locality, causality, equivalence principle.
• Conservation Laws: Energy, momentum, lepton/baryon number.
• Regularity Properties: Renormalizability, unitarity, analyticity.

Crucially, the grammar is distinct from the specific Lagrangian; it represents the regularities that survive across reformulations and limit cases.

2.2 Coherence Cost and Prior

For a candidate model $M$, the method calculates a coherence cost $C(M|T)$:
$$C(M|T) = \sum_{i=1}^n w_i \delta_i(M, T)$$
where:

• $\delta_i(M, T) = 1$ if $M$ violates principle $p_i$ without independent motivation, and $0$ otherwise.
• $w_i \geq 0$ are optional weights reflecting the empirical robustness of the principle (default $w_i=1$).
• "Independent motivation" must be empirical or theoretical, not aesthetic.

The coherence prior is then defined via a maximum-entropy exponential form:
$$P(M|T) \propto \exp[-\alpha C(M|T)]$$
where $\alpha > 0$ is a calibratable scale parameter. In the minimal implementation, $\alpha = 1$, assigning a penalty factor of $e^{-1} \approx 0.37$ for each unmotivated violation.

2.3 Distinction from Existing Methods

The authors explicitly distinguish the Coherence Principle from:

• The Occam Factor: The Occam factor (internal to the Bayes factor) penalizes unused parameter space volume. The Coherence Principle operates at the model space level, penalizing structural disagreement with a paradigm before data analysis.
• Reference Priors: Reference priors are designed to be agnostic to external theoretical knowledge. The Coherence Principle explicitly incorporates such knowledge, which is crucial when data are insufficient.
• Naturalness: Naturalness arguments are often scale-dependent and have counterexamples. The Coherence Principle relies on explicitly listed, falsifiable structural rules rather than a single quantitative scale.

3. Key Applications and Results

The paper illustrates the principle across several domains, demonstrating that it reproduces qualitative rankings already used informally in the field but makes them quantitative and reproducible.

3.1 Neutrino Mass Mechanisms

Comparing a Unified Model (family-universal extension, e.g., Type-I seesaw) vs. a Disjoint Model (different mechanisms for different families).

• Result: The Disjoint model violates family universality ($C=1$), while the Unified model respects the grammar ($C=0$).
• Impact: With $\alpha=1$, the prior odds favor the Unified model by a factor of $e \approx 2.72$. Combined with existing data, this shifts the evidence from "anecdotal" to "positive" for the normal ordering, resolving an ambiguity that depends heavily on prior construction.

3.2 Dark Energy and Modified Gravity

The grammar of General Relativity ($G(T_{GR})$) includes principles like general covariance, the equivalence principle, and second-order field equations.

• Results:
  • Cosmological Constant ($\Lambda$) & Minimal Quintessence: Respect all rules ($C=0$).
  • Screened Scalar-Tensor ($f(R)$, Horndeski): Often require screening mechanisms to avoid equivalence principle violations; if the screening is not self-consistent, $C \approx 1$.
  • Lorentz-Violating/Nonlocal Gravity: Violate locality or Lorentz invariance ($C \approx 2-3$).
• DESI Preference: The DESI data hints at a phantom crossing ($w < -1$). A single minimally coupled scalar cannot cross the phantom divide without instability (Vikman's no-go). Realizing this requires extra degrees of freedom or modified gravity, incurring a coherence cost ($C \geq 1$). The prior penalizes the crossing realization, suggesting the data favors late-time dynamics rather than a genuine phantom crossing.

3.3 Inflation and Beyond-Standard-Model (BSM)

• Inflation: Single-field slow-roll models ($C=0$) are favored over models with detectable isocurvature or non-Bunch-Davies initial states ($C=1$).
• BSM: Universal extensions (e.g., MSSM, GUTs) have $C=0$. Non-universal gauge couplings (e.g., family-dependent $Z'$) violate family universality ($C=1$) unless independently motivated by anomalies.

3.4 Historical Case Studies (Retrodiction)

The authors test the principle on historical paradigm shifts to verify if it favors the historically successful choice when applied with the correct, time-stamped grammar.

• Einstein vs. Newton:
  • Pre-1905 Grammar: Based on Newtonian mechanics. General Relativity (GR) violates absolute time and Galilean invariance ($C \approx 4$). The principle would have disfavored GR.
  • Post-1905 Grammar: Based on Special Relativity (validated by 1915). Newtonian gravity violates Lorentz invariance ($C \geq 2$), while GR respects it ($C \approx 0$). The principle correctly favors GR.
  • Lesson: The grammar must be time-stamped to the most secure currently validated theory.
• Pauli's Neutrino vs. Abandoning Conservation:
  • In 1930, data could not distinguish between a new particle (neutrino) and statistical energy conservation.
  • The "Neutrino" model respects all conservation laws ($C=0$). The "Bohr" model violates energy, momentum, and angular momentum ($C \approx 3$).
  • The principle strongly favors Pauli ($e^3 \approx 20$), correctly predicting the historical outcome despite the lack of direct detection at the time.
• Parity Violation:
  • A naive application using a grammar that included parity conservation as a universal rule would have disfavored Lee and Yang's proposal.
  • The resolution: Parity conservation was only validated in strong/electromagnetic sectors, not weak. A domain-resolved grammar assigns no cost to parity violation in the weak sector. The "failure" was an over-extended grammar, not the principle itself.
• Special Relativity vs. Lorentz Aether:
  • In 1905, both theories were empirically degenerate ($B \approx 1$).
  • The Aether theory retains a preferred frame (violating the principle of relativity, $C=1$). Special Relativity elevates the principle to a postulate ($C=0$).
  • The prior alone favors Special Relativity by $e^1 \approx 2.7$, acting as the decisive factor in a degenerate regime.

3.5 Signed Extension: The Coherence Bonus

The authors propose a signed extension where a model can earn a "coherence bonus" (negative cost) if it explains a principle that was previously an unexplained coincidence in the background theory (e.g., GR deriving the equivalence principle rather than assuming it). This sharpens the prior odds further (e.g., favoring GR over Newton by $e^3 \approx 20$).

4. Significance and Claims

The paper claims the Coherence Principle offers a reproducible, falsifiable, and transparent method for incorporating theoretical structural knowledge into Bayesian inference.

• Reproducibility: Every input (the grammar $G(T)$, weights, and indicators) is explicit and open to challenge.
• Falsifiability: The prior odds are finite and can be overturned by sufficiently strong data.
• Modesty: The authors emphasize that the principle is not a replacement for the likelihood, nor a double-counting of the Occam factor. It does not rely on aesthetic criteria (elegance/beauty) unless they can be restated as empirically validated structural rules.
• Scope: While illustrated in cosmology and particle physics, the framework is presented as a general method for theory-informed inference applicable to any field where a validated theoretical framework constrains the space of admissible extensions (e.g., biology, economics, climate science).

The paper concludes that the Coherence Principle converts the informal "trust in validated structural rules" used by physicists into a routine, quantitative component of Bayesian model selection, particularly valuable in regimes where data are weak or degenerate.