Needs Verification

Analogical Support

Introduction & Epistemological Purpose¶

In the philosophy of science and automated theory analysis, analogical boost describes the process by which a novel hypothesis \(H_1\) in a specific field (the target domain \(D_1\)) gains credibility and epistemic weight by mirroring the mathematical or functional structure of an already established, proven hypothesis \(H_2\) from another field (the source domain \(D_2\)). Classic historical examples include the transfer of model structures from hydrodynamics (water waves) to acoustics (sound waves) or optics (light waves).

The metric of the Analogical Boost (Lateral Analogical Boost, \(LAB_{\text{norm}}\)) quantifies this excitatory (supporting) effect with mathematical rigor. It combines methods from network theory (regular equivalence on bounded neighborhoods) with principles from cognitive philosophy of science (Thagard's Explanatory Coherence).

Graph-Theoretical Foundations¶

To process theories and hypotheses mathematically, scientific knowledge is modeled as a directed or undirected theory graph \(G = (V, E)\):

Nodes \(V\): Represent individual hypotheses, law-like statements, or theory elements \(H \in V\). Each node is assigned to a specific scientific domain \(\text{Dom} (H)\).
Edges \(E\): Represent relational dependencies, derivations, or structural interactions between the hypotheses.

The Concept of Regular Equivalence¶

Genuine scientific analogies are characterized by the fact that two hypotheses from different fields share hardly any common technical terms or direct connecting edges, but occupy the same relational roles in their respective subnetworks.

In network theory, two types of node similarities are distinguished:

Structural Equivalence: Two nodes share the same direct neighbor nodes. (Equivalent to synonymy within the same domain).
Regular Equivalence: Two nodes \(H_1\) and \(H_2\) are similar if they have neighbor nodes that are themselves similarly structured – even if \(H_1\) and \(H_2\) do not share a single common neighbor. It is precisely this measure that captures abstract, cross-domain analogies.

The \(k\)-Hop Neighborhood¶

To prevent distant, irrelevant paths in the theory graph from distorting the similarity measure or slowing down the calculation, the metric locally restricts the structural analysis to a \(k\)-hop neighborhood. This means that only relationship paths of maximum length \(k\) (edge steps) are considered.

Mathematical Formulation of the Metric¶

The calculation is performed in three consecutive steps: (1) determination of structural-relational similarity, (2) calculation of the cross-domain weight, and (3) saturation of the total effect.

Step 1: Bounded Regular Equivalence (\(\text{Sim}_{\text{reg}}^{ (k)}\))¶

Let A \(\mathbf{A}\) be the adjacency matrix, and \(\mathbf{D}\) the degree diagonal matrix where the \(i\)-th diagonal entry \(D_{ii} = \sum_{j} A_{ij}\) corresponds to the total number (or sum of weights) of all outgoing edges from node \(i\). By multiplying with the inverse degree matrix \(\mathbf{D}^{-1}\), the rows of \(\mathbf{A}\) are scaled so that their entries provide information about the relative connection proportions (transition probabilities on the graph). Let \((\cdot)^r\) be the matrix power, where he \(r\)-th power of the normalized adjacency matrix yields the sum of the weighted paths of exact length \(r\) between two nodes.

The relational similarity of two hypotheses \(H_1\) and \(H_2\) within a \(k\)-hop environment is calculated via a damped and degree-normalized path matrix:

\[\text{Sim}_{\text{reg}}^{ (k)} (H_1, H_2) = \sum_{r=1}^{k} \alpha^r \cdot \left ( (\mathbf{D}^{-1}\mathbf{A})^r \right)_{H_1, H_2}\]

Explanation of the mathematical terms in this formula¶

\(r\) (Path Length): The running index of the sum, which incrementally adds up paths from length \(1\) to a maximum of \(k\).
\(k\) (Search Depth / \(k\)-Hop Variable): A freely selectable integer control variable (\(k \in \mathbb{N}^+\)) that specifies how far the local network environment should be analyzed.
Default Value: \(k_{\text{default}} = 3\) (captures direct connections, 2-edge motifs, and 3-edge contextual structures).
\(\alpha\) (Katz Damping Factor): A scalar in the range \(0 < \alpha < 1\) that exponentially devalues longer paths (\(\alpha^r\)), since distant relational connections possess less epistemic significance.

Step 2: Cross-Domain Weight Function (\(w_{\text{analogy}}\))¶

To ensure that the metric specifically evaluates analogies between different fields, the structural similarity is combined with a domain filter:

\[w_{\text{analogy}} (H_1, H_2) = w_{\text{base}} \cdot \phi_{\text{domain}} (H_1, H_2) \cdot \text{Sim}_{\text{reg}}^{ (k)} (H_1, H_2)\]

(Where \(\phi_{\text{domain}} (H_1, H_2)\) is the Cross-Domain Indicator Function: A binary control function that checks whether the two hypotheses originate from different domains:

\[\phi_{\text{domain}} (H_1, H_2) = \begin{cases} 1 & \text{if } \text{Dom} (H_1) \neq \text{Dom} (H_2) \\ 0 & \text{if } \text{Dom} (H_1) = \text{Dom} (H_2) \end{cases}\]

This guarantees that pure intra-domain similarities are not falsely evaluated as analogies.

Property of Non-Negativity (\(w_{\text{analogy}} \ge 0\)): Since all components of the formula are non-negative, \(w_{\text{analogy}} \ge 0\) strictly applies. This guarantees that analogies always form excitatory (supporting) connections and never falsely generate inhibitory (contradictory) edges).

Step 3: Normalized Lateral Analogical Boost (\(LAB_{\text{norm}}\))¶

The total boost that the target hypothesis \(H_1\) receives from all active analogous source hypotheses \(H_2\) in the network is calculated through a saturated summation over all existing nodes, excluding \(H_1\) itself. This accumulation scales the analogical weight by \(a_{H_2} (t) \in [-1, 1]\) representing the overall corroboration of the source hypothesis at time \(t\) (where corroborated hypotheses, \(a_{H_2} (t) > 0\), primarily drive the boost). Finally, the hyperbolic tangent (\(\tanh\)) is applied as a saturation function mapping the input onto the interval \([0, 1)\), which prevents the target hypothesis from accumulating infinitely high activation values through the aggregation of many weak analogies and ensures mathematical stability and convergence during network relaxation.

\[LAB_{\text{norm}} (H_1) = \tanh \left (\sum_{H_2 \in V \setminus \{H_1\}} w_{\text{analogy}} (H_1, H_2) \cdot a_{H_2} (t) \right)\]

Parameter and Variable Overview¶

Symbol	Name / Meaning	Mathematical Type / Range	Default Value
\(k\)	\(k\)-hop search depth	Positive integer (\(\mathbb{N}^+\))	\(3\)
\(\alpha\)	Katz damping factor	Real number in \((0, 1)\)	\(0.1\)
\(w_{\text{base}}\)	Base connection weight	Real number in \((0, 1]\)	\(0.2\)
\(\mathbf{A}\)	Adjacency matrix of the theory graph	\(N \times N\) matrix with \(A_{ij} \ge 0\)	From graph structure
\(\mathbf{D}\)	Degree diagonal matrix	\(N \times N\) diagonal matrix (\(D_{ii} = \sum_j A_{ij}\))	From matrix \(\mathbf{A}\)
\(\phi_{\text{domain}}\)	Domain filter indicator	Binary value \(\{0, 1\}\)	\(1\) for \(D_1 \neq D_2\)
\(a_{H_2} (t)\)	Activation of the source hypothesis	Real number in \([-1, 1]\)	Dynamic from network
\(LAB_{\text{norm}} (H_1)\)	Normalized analogical boost	Real number in the interval \([0, 1)\)	Calculated output

Literature & References¶

Balzer, W., Moulines, C. U., & Sneed, J. D. (1987): An Architectonic for Science: The Structuralist Program. Reidel Publishing Company.
Newman, M. (2018): Networks: An Introduction (2nd ed.). Oxford University Press.
Schurz, G. (2014/2024): Philosophy of Science: A Unified Approach. Routledge.
Stegmüller, W. (1976): The Structure and Dynamics of Theories. Springer-Verlag.
Thagard, P. (1989): Explanatory Coherence. Behavioral and Brain Sciences, 12 (3), 435–502.