Confidence-Evidence Ratio (Ψ)

• Confidence-Evidence Ratio (Ψ) is a metric that quantifies the relationship between subjective confidence and additive evidence in probabilistic inference frameworks.
• It is derived from Bayesian and axiomatic learning principles, converting fractional confidence into additive evidence using log-based transformations with context-dependent parameters.
• The ratio is pivotal in assessing dataset consistency, informing cosmological tension tests, and unifying approaches from Kalman filtering to Dempster–Shafer theory.

The Confidence-Evidence Ratio, denoted $\Psi$ (and alias $R$ in certain contexts), formalizes the relationship between a quantitative measure of confidence and an associated amount of evidence in probabilistic inference and belief-updating frameworks. It arises from both Bayesian statistics and axiomatic theories of learning with confidence, where it links the multiplicative structure of belief updates to an additive quantification of evidence or trust. $\Psi$ serves both as a diagnostic in model comparison—especially for data-set consistency/tension in cosmology—and as a foundational object in abstract learning frameworks, mediating the flow from subjective confidence measures to structural update dynamics.

1. Definitions and Mathematical Formulation

In Bayesian inference, the Confidence-Evidence Ratio $\Psi$ offers a probabilistically meaningful measure of how much "extra" confidence one should place in the joint compatibility of two datasets. Given Bayesian evidences $\mathcal{Z}_A$, $\mathcal{Z}_B$ for datasets $A$ and $B$, and their joint evidence $\mathcal{Z}_{AB}$ under a shared prior $\pi(\theta)$, the ratio is defined as: $R$0 This admits several equivalent formulations, notably in terms of posterior distributions $R$1: $R$2 In axiomatically-motivated frameworks, such as learning with confidence, $R$3 emerges as a conversion parameter between fractional confidence $R$4 and additive evidence $R$5: $R$6 The parameter $R$7 is the Confidence-Evidence Ratio and is context-dependent, typically taking the value 1 in standard Bayesian updates and Kalman filtering settings (Richardson, 14 Aug 2025).

2. Axiomatic Foundations and Canonical Domains

A rigorous foundation for $R$8 arises from a set of axioms for confidence-based updates, central to the work "Learning with Confidence" (Richardson, 14 Aug 2025). The relevant space is characterized as follows:

• $R$9: space of belief states (priors)
• $\Psi$0: space of possible observations
• $\Psi$1: domain of confidence values, supporting binary associative operation $\Psi$2 (composition), partial order, and identity elements $\Psi$3 (no confidence) and $\Psi$4 (full confidence).

The axioms CF0–CF4 (zero confidence, idempotence, smoothness, associativity, and non-cyclicity) guarantee that any such update admits a reparameterization into an additive flow: $\Psi$5 For the canonical continuous domains—(a) fractional $\Psi$6 with $\Psi$7-combination, and (b) additive $\Psi$8 with ordinary addition—an isomorphism exists via $\Psi$9, with $\Psi$0 as a scale parameter: $\Psi$1 $\Psi$2 acts as the conversion factor between fractional trust and linearized (additive) evidence.

3. Prior Dependence and the Role of KL Divergences

A salient property of the Bayesian version of $\Psi$3 is its explicit dependence on the prior volume $\Psi$4. As shown in "Quantifying tensions in cosmological parameters" (Handley et al., 2019), for uniform (top-hat) priors,

$\Psi$5

and for Gaussian posteriors within a box prior,

$\Psi$6

Thus, shrinking the prior volume always drives down $\Psi$7 (and $\Psi$8), which can artificially mask incompatibility between datasets. The recommendation is that if any physically reasonable prior width yields $\Psi$9, the datasets should be deemed to be in tension.

To correct for this, the prior-independent modification uses Kullback–Leibler divergences $\mathcal{Z}_A$0: $\mathcal{Z}_A$1 The suspiciousness statistic $\mathcal{Z}_A$2 is then defined as: $\mathcal{Z}_A$3 $\mathcal{Z}_A$4 isolates genuine model tension, free of prior volume artifacts, and supports calibrated statistical interpretation.

4. Interpretations in Learning and Inference

Across learning theory and Bayesian updating, $\mathcal{Z}_A$5 governs how different types of confidence feedback map onto additive evidence accumulation. In belief flows induced by learning rules,

• The vector-field representation interprets the path $\mathcal{Z}_A$6 as an integral curve.
• In a loss-based (gradient) model, the relationship between infinitesimal update rates and total evidence follows directly from $\mathcal{Z}_A$7: $\mathcal{Z}_A$8 In the Bayesian case, with $\mathcal{Z}_A$9, the soft-update via likelihood ratio $\mathcal{Z}_B$0 raised to the $\mathcal{Z}_B$1-th power recovers ordinary Bayes rule for $\mathcal{Z}_B$2. For Kalman filtering, the confidence-evidence mapping reproduces the Kalman gain structure, with

$\mathcal{Z}_B$3

A similar structure incorporates Shafer's weight of evidence and Dempster–Shafer theory, reinforcing the centrality of $\mathcal{Z}_B$4 (Richardson, 14 Aug 2025).

5. Practical Computation and Statistical Usage

The computation of $\mathcal{Z}_B$5 in data-set cross-consistency testing is rigorous and systematic (Handley et al., 2019). The recommended workflow is:

1. Compute the raw ratio $\mathcal{Z}_B$6 using the joint and marginal evidences for the datasets under consistent priors.
2. Explore the dependence of $\mathcal{Z}_B$7 on prior width; search for any physically reasonable priors with $\mathcal{Z}_B$8 as a necessary test for tension.
3. Calculate $\mathcal{Z}_B$9 for a prior-invariant assessment, and report the tension probability $A$0 based on a calibrated $A$1-distribution:

$A$2

where $A$3 is the effective dimension (from Bayesian model dimensionality).

The table below summarizes the qualitative interpretation:

$A$4/$A$5	$A$6 value	Tension classification
$A$7, $A$8	$A$9	No tension
Some $B$0, $B$1	$B$2	Moderate tension
All $B$3, $B$4	$B$5	Strong tension

6. Special Cases and Connections

• Cosmological parameter tension: $B$6 and variants are routinely applied to test compatibility between major data sets (e.g., Planck, DES, BOSS, SHOES). Empirical results indicate graduated degrees of tension depending on the datasets and priors, supporting fine-grained model assessment (Handley et al., 2019).
• Kalman filter and Dempster–Shafer weight: In dynamic and epistemic uncertainty models, $B$7 recovers classical weights of evidence and trust scaling, unifying several learning paradigms under a shared conversion law (Richardson, 14 Aug 2025).

7. Significance and Recommended Practice

$B$8 provides a mathematically principled and operationally effective interface between subjective trust, additive evidence, and model/data consistency. Correct usage requires careful attention to the prior dependence (in Bayesian settings) and the correct interpretation of $B$9 in abstract learning scenarios. Reporting both the raw Bayesian confidence ratio $\mathcal{Z}_{AB}$0 and its prior-independent counterpart $\mathcal{Z}_{AB}$1 (with effective dimension $\mathcal{Z}_{AB}$2) is advocated for transparency and reproducibility, especially in inference problems involving high-stakes scientific consistency tests. This dual reporting framework is now a proposed standard in cosmological data analysis (Handley et al., 2019), and the axiomatic translation underscores its universality in learning and evidence aggregation (Richardson, 14 Aug 2025).