Adherence-Structure Balance (ASB)

• Adherence-Structure Balance (ASB) is a framework that quantifies how individual rule adherence induces overall structural patterns in networks.
• It applies multi-scale indices—from triad-level metrics to global frustration measures—to diagnose polarization and faction formation in complex systems.
• ASB spans domains such as statistical mechanics, causal inference, and expert-driven optimization, highlighting phase transitions and separable effects in decision processes.

Adherence-Structure Balance (ASB) captures the nuanced interplay between the adherence of individual agents or components to prescribed rules or recommendations and the emergence or sustenance of large-scale structural patterns (such as faction formation, polarization, or global system balance). The concept arises across domains—from signed social networks and statistical mechanics to causal inference in adherence-aware interventions and expert-in-the-loop decision optimization. ASB frameworks integrate local compliance (e.g., norm-following or adherence to a recommendation) with meso- and macro-level organizational structure, offering both empirical diagnostics and theoretical insight into collective phenomena.

1. Formalism and Multilevel Profile in Signed Networks

In directed signed networks, ASB is concretized via a three-tiered analytic protocol grounded in balance theory (Aref et al., 2020). The system is a directed graph with edge signs $\sigma_{ij}\in\{+1,-1\}$ (friend/enemy). ASB is quantified by aligning adherence to local balance axioms with network-scale structural polarization, operationalized at micro, meso, and macro levels:

• Triad-Level (Micro): Focuses on tension distribution in each triple of nodes, emphasizing transitive semicycles—directed 3-cycles where the adjacency’s transitivity is honored. The triad-level balance index is

$$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$

where $\mathcal{T}$ is the set of all transitive semicycles in $G$.

• Subgroup-Level (Meso): Identifies whether nodes can be partitioned into two internally cohesive, mutually antagonistic camps, evaluated by cohesiveness $C^*$ (fraction of intra-camp positive ties) and divisiveness $D^*$ (fraction of inter-camp negative ties) with respect to the optimal bipartition that minimizes frustrated edges.
• Network-Level (Macro): Employs the normalized line index

$F(G) = 1 - \frac{2L(G)}{m}$

where $L(G)$ is the frustration index (minimum edges to flip or remove for global balance), and $m$ is the total number of edges.

This multiscale approach yields a diagnostic ASB profile $\{T(G), C^*, D^*, F(G)\}$, enabling comparative analysis across empirical systems such as online communities, tribal alliances, and scholarly networks. The simultaneous assessment at all three scales disambiguates contexts where high local adherence may not produce global structural polarization, or vice versa (Aref et al., 2020).

2. Statistical Mechanics: Competing Orders in Networks

ASB also admits a statistical-physics formalization in models that interpolate between adherence-driven (local, coevolutionary) balance and structure-driven (global, Heider-type) balance (Noudehi et al., 2022). Here, every node $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$0 carries a spin $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$1 and links $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$2. Two Hamiltonians encapsulate the energetic preferences:

• Coevolutionary (Adherence) Hamiltonian:

$$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$3

Favors node-link-node agreement.

• Structural (Global) Hamiltonian:

$$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$4

Favors triangles with positive signature.

The total Hamiltonian combines these: $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$5 for $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$6 interpolating between adherence and structural dominance. Mean-field self-consistency yields order parameters (magnetizations and correlations) whose bifurcations define phase transitions:

• Continuous (second-order) transitions at small $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$7 (adherence-dominated);
• Discontinuous (first-order) transitions for large $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$8 (structure-dominated);
• A tricritical point separates these regimes, with Monte Carlo simulations corroborating mean-field predictions.

This formalism reveals that under varying environmental "temperature" and competition between local and global rules, system-level balance can arise through continuous or abrupt collective shifts—mapping onto real social phenomena like sudden polarization or gradual consensus (Noudehi et al., 2022).

3. ASB in Causal Inference: Separable Effects for Adherence

In causal inference, ASB arises in the context of defining estimands that isolate effects of interest while neutralizing adherence-related confounding (Wanis et al., 2023). Partitioning a binary treatment $$T(G) = \frac{|\{t\in\mathcal{T}: \text{sign}(t)=+1\}|}{|\mathcal{T}|}$$9 into $\mathcal{T}$0 (adherence-causing) and $\mathcal{T}$1 (active component), a separable effect compares outcomes under interventions that vary $\mathcal{T}$2 while holding $\mathcal{T}$3 fixed. Formally: $\mathcal{T}$4 Under assumptions encoded in a DAG (including $\mathcal{T}$5, no unmeasured confounding beyond $\mathcal{T}$6), adherence is structurally balanced: both arms yield identical adherence distributions, and observed effects on the outcome are disentangled from adherence artifacts. Semi-parametric estimators (e.g., IPW, g-formula) enable identification under these constraints. This contrasts sharply with usual intention-to-treat or per-protocol estimands, positioning separable effects as an adherence–structure balance solution in comparative effectiveness research (Wanis et al., 2023).

4. ASB in Expert-in-the-Loop Optimization

High-stakes decisions, such as medical recommendations, require optimization schemes that are robust to human-in-the-loop deviations. The adherence-aware optimization framework models the effective policy as a convex combination of the algorithmic recommendation $\mathcal{T}$7 and a baseline $\mathcal{T}$8 weighted by adherence parameter $\mathcal{T}$9: $G$0 This formulation leads to an optimization problem (ASB-MDP) maximizing actual implementation performance. The solution structure exhibits:

• Existence of simple (stationary, deterministic) optimal $G$1,
• Monotonic and piecewise-constant dependence of performance on $G$2,
• Robustness to uncertainty in $G$3 and hard constraints on states of adherence,
• Efficient resolution via value iteration, policy iteration, or linear programming (Grand-Clément et al., 2022).

Empirical results indicate substantial performance gains when $G$4 is recalibrated for realistic adherence rates, confirming that the ASB perspective closes the gap between prescriptive and effective policies.

5. Dynamical Realization: Indirect Reciprocity and Weak Balance

ASB also emerges naturally from the distributed dynamics of indirect reciprocity as a generator of weakly structurally balanced states in social networks (Bae et al., 10 Jan 2025). Under the "judging" norm, each individual's cooperative behavior is norm-driven and assessments are updated based on a third-order rule. Stationary configurations align precisely with weak structural balance—no cycles with exactly one negative edge.

• Micro level: Agents adhere to the norm, leading to fast stabilization into a weakly balanced partition.
• Meso/Macro level: Occasional assessment errors induce transitions, shaping the long-run cluster size distribution: a “giant” cluster persists alongside a heavy-tailed ensemble of smaller factions. The micro-level adherence realizes mesoscopic (factional) and macroscopic (polarization profile) structure.
• The process exhibits a distinction in timescales: rapid local norm conformity induces structural order, while rare errors modulate global structure.

The same multilevel logic of ASB applies: local adherence ensures compatibility with balance constraints, while system-level structure emerges from the interaction between adherence and stochastic perturbations (Bae et al., 10 Jan 2025).

6. Unifying Themes and Distinctions

Across these domains, the following patterns characterize ASB:

Setting	Adherence Mechanism	Structure Realized
Signed Social Networks (Aref et al., 2020)	Obedience to balance axioms	Polarized factions, network-wide balance
Statistical-Physics (Noudehi et al., 2022)	Local energy minimization	Phase transitions, order-disorder
Causal Inference (Wanis et al., 2023)	Fixing adherence-causing component	Isolated treatment effects with balanced adherence
Expert-in-the-Loop (Grand-Clément et al., 2022)	Partial compliance with policy	Realistic, robust optimality
Indirect Reciprocity (Bae et al., 10 Jan 2025)	Norm-driven updates with errors	Self-organized factional structure

The general principle is that adherence at the local or component level under specified constraints or protocols induces structure at higher scales. The specific techniques, indices, or optimization formulations operationalize this general principle according to domain requirements.

A plausible implication is that the robustness and expressiveness of systems exhibiting ASB hinge on the proper calibration of local adherence rules and their consistency with global structuring objectives. Moreover, phase transition phenomena and the diagnostic separation of multi-scale indices underscore the complexity and richness of real-world systems where adherence and structure are intertwined.