Independent Approval Model (IAM)

• IAM is a framework that decomposes complex interactions into independent probabilistic or physical events, enabling tractable mathematical analysis and efficient estimation.
• In voting theory, IAM models voter approvals via independent Bernoulli trials, providing insights into justified representation and other computational challenges.
• IAM underpins structured regulatory pipelines for frontier AI systems and models molecular collision events using the Independent Atom Model with geometrical corrections.

The Independent Approval Model (IAM) encompasses several influential frameworks unified by the assumption of statistical or procedural independence applied to approval, voting, regulation, or physical interactions. IAM spans domains including social choice under uncertainty, generative modeling for voting data, regulatory approval of high-risk systems, and molecular collision physics. Despite the diversity of contexts, all IAM instances enforce a decomposition into elementary, independent probabilistic or physical events, facilitating tractable mathematical analysis and often permitting efficient parameter estimation or inference. This article provides an integrated exposition of the IAM as rigorously formalized in approval-based voting theory, generative models for approval elections, regulatory processes for AI systems, and collision physics.

1. Formal Definitions and Domains of Application

1.1 Voting Theory and Social Choice

In approval-based committee voting with uncertainty, the Independent Approval Model (also termed the Candidate-Probability model) is formulated as follows (Aziz et al., 2024):

Given voter set $V$ ($|V|=n$) and candidate set $C$ ($|C|=m$), each pair $(v,c)$ is associated with a parameter $p_{v,c} \in [0,1]$, denoting the probability that $v$ approves $c$. Each voter's approval for each candidate is realized as an independent Bernoulli trial. The complete approval profile $A: V \times C \to \{0,1\}$ occurs with probability

$$\Pr[A] = \prod_{v \in V} \prod_{c \in C} p_{v,c}^{A(v,c)} \cdot (1 - p_{v,c})^{1 - A(v,c)}.$$

1.2 Probabilistic Generative Models for Election Data

IAM generalizes models such as impartial culture and Bernoulli resampling in the generation of approval ballots (Faliszewski et al., 26 Jan 2026). For $n$ candidates, each with an approval probability $p_i$, a voter's approval set $A$ is generated by independent Bernoulli draws:

$$P(A) = \prod_{i \in A} p_i \prod_{i \notin A} (1 - p_i).$$

1.3 Regulatory Approval of AI Systems

In regulatory frameworks for “frontier AI,” IAM refers to regimes in which explicit, independent regulatory authority must approve critical system transitions (training and deployment) based on safety evidence (Salvador, 2024). Here, “independent approval” is institutional rather than probabilistic; regulatory decisions follow a staged, auditable pipeline structurally paralleling the statistical independence in other IAM domains.

1.4 Collision Physics: The Independent Atom Model

In molecular collision physics, the Independent Atom Model postulates that ionization, capture, or related reactions proceed independently on each atomic center (Lüdde et al., 11 Sep 2025). Cross sections are summed or geometrically screened (using the pixel-counting method) under the assumption of inter-atomic independence.

2. Mathematical Structure and Statistical Properties

2.1 Independence Assumptions

IAM enforces:

• Voter–voter independence: Approval sets of distinct voters are independent.
• Within-voter, candidate–candidate independence: Each approval event for a given voter is independent across candidates.
• Physical independence: In collision models, atomic centers interact independently with a projectile, neglecting inter-atomic quantum coherence (Lüdde et al., 11 Sep 2025).

These independence assumptions define the joint distribution as a product of marginals.

2.2 Uncertainty Representation

By parameterizing $p_{v,c}$ (or $p_i$ in single-population models), IAM encodes uncertainty either empirically (from past data), structurally (as model parameters), or operationally (as probabilistic beliefs or group-averaged frequencies). In generative models, these probabilities are estimated via maximum likelihood or Bayesian inference, and may be extended to mixture models to capture heterogeneity (Faliszewski et al., 26 Jan 2026).

3. Computational Questions and Algorithmic Results

Under IAM, several fundamental computational questions arise in voting theory (Aziz et al., 2024):

• IsPossJR ("Possible JR"): In polynomial time, one can decide whether a given committee $W$ satisfies justified representation (JR) with nonzero probability.
• IsNecJR ("Necessary JR"): In polynomial time, one can decide whether $W$ satisfies JR with probability one.
• ExistsNecJR: Determining existence of any necessarily JR-satisfying committee is NP-complete.
• JR-Probability: Computing the exact probability that $W$ satisfies JR is #P-complete.
• MaxJR: Optimizing probability of JR satisfaction is NP-hard.

Similar complexity classifications extend to Proportional Justified Representation (PJR) and Extended Justified Representation (EJR), with coNP-completeness and coNP-hardness for corresponding decision and optimization problems.

4. Learning, Estimation, and Mixture Extensions

4.1 Maximum Likelihood and Bayesian Estimation

Given observed approval ballots $A_1,\dots,A_m$, the MLE for each $p_i$ is the observed approval frequency:

$p_i^* = \frac{1}{m} \sum_{j=1}^{m} 1_{i \in A_j}$

Bayesian estimation proceeds via independent Beta priors, yielding analytically tractable posteriors and closed-form posterior means for $p_i$ (Faliszewski et al., 26 Jan 2026).

4.2 Mixture IAMs

Single IAMs often fail to capture multimodal or heavy-tailed approval patterns visible in real-world election data. Mixtures of IAMs, learned via Expectation-Maximization (EM) or Bayesian inference, assign each vote to one of $K$ components (with per-component parameters $\{p_{k,i}\}$ and weights $\alpha_k$), enabling realistic generative modeling. Mixtures nearly halve the discrepancy (measured by normalized Hamming distance) between synthetic and real votes relative to single IAMs (Faliszewski et al., 26 Jan 2026).

5. Regulatory IAM: Approval Pipelines for Frontier AI

IAM, as an approval regulation regime for AI, establishes:

• Compute-based gating: Models are covered if cumulative training FLOPs meets/exceeds $C^* = 10^{26}$.
• Two approval gates: Training Authorization (with upfront security, capability, and compute documentation) and Model Certification prior to deployment.
• Compliance protocols: Certification basis, project-specific plans, and model deployment cards structure evaluation and enforcement.
• Continuous monitoring: Post-deployment, incident rates and periodic capability evaluations trigger possible model withdrawal or review (Salvador, 2024).

Significant challenges in implementation include enforcement against covert deployments, specification of comprehensive certification requirements, experimental reliability, pre-filtering of safe models, and regulatory overhead. Adjustable compute thresholds, routine audits, and public regulatory code development are recommended mitigations.

6. Physical IAM: Molecular Collision, PCM, and xPCM

The Independent Atom Model in collision physics constructs the molecular cross section as the sum over atomic cross sections, corrected for geometric screening using the pixel-counting method (PCM). PCM computes the effective area by overlaying 2D atomic disks and averaging their union area over molecular orientations. The generalized xPCM incorporates the possibility of multiple sequential in-molecule collisions, introducing corrections based on mean free paths ($\lambda_i$) and inter-atomic distances (Lüdde et al., 11 Sep 2025). For low-charge projectiles or small molecules, PCM and xPCM yield similar cross sections. Double-scattering effects become significant for highly charged projectiles or larger molecules in certain energy regimes.

7. Illustrative Examples and Open Problems

A worked example in approval voting considers $V = \{v_1, v_2\}$, $C = \{a, b, c\}$, $k = 1$, with specified $p_{v,c}$. The probability of a designated approval profile and the computation of JR probability (over $2^{n m}$ profiles) are demonstrated. For small cases, enumeration is feasible; in general, such calculations are #P-hard (Aziz et al., 2024).

Open problems include development of approximation or fixed-parameter tractable algorithms for MaxJR, mechanisms to refine or query $p_{v,c}$, extension of IAM to other objectives, empirical analysis of IAM-based voting rules, and strategic reporting of uncertain approvals. In physical applications, ongoing refinement of mean-free-path corrections and screening algorithms is necessary as new collision regimes are explored.

---

References:

(Aziz et al., 2024, Faliszewski et al., 26 Jan 2026, Salvador, 2024, Lüdde et al., 11 Sep 2025)