Geometric Evidential Head

• Geometric evidential head is a modeling component that assigns belief measures using explicit geometric rules and deterministic sensor models.
• It utilizes mathematical fusion rules such as Dempster’s and Yager’s to create sharp, interpretable state boundaries and ensure convergence in occupancy mapping.
• The approach integrates deep, prior-driven evidence with geometric constraints, enhancing robustness in autonomous navigation and protein structure prediction.

A geometric evidential head is a rigorously specified modeling component that assigns and propagates belief measures (evidence masses) over discrete states—such as occupancy or prediction uncertainty—based directly on geometric rules, priors, and/or deterministic sensor models, rather than learning from data. This approach is used in domains ranging from evidential occupancy mapping in autonomous vehicle perception (Bauer et al., 2020) to robust uncertainty quantification in protein structure prediction (Shihab et al., 12 Jan 2026). Its defining features include explicit mathematical assignment of evidential parameters, fusion via formal combination rules, and convergence properties guaranteed by the underlying geometry, which together yield sharp, interpretable boundaries, stable uncertainty under distribution shift, and controlled integration with learned or prior-driven components.

1. Mathematical Foundations of Geometric Evidential Heads

The geometric evidential head rests on explicit probabilistic or evidential assignments dictated by sensor geometry or spatial constraints. In evidential occupancy grids for radar-based mapping, each cell $(i,j)$ is assigned a mass vector

$$m^{(k)}_{ij} = \begin{bmatrix} m^{(k)}_{f,ij} \ m^{(k)}_{o,ij} \ m^{(k)}_{u,ij} \end{bmatrix},\quad m^{(k)}_{f,ij} + m^{(k)}_{o,ij} + m^{(k)}_{u,ij} = 1$$

where the entries correspond to evidence for “free”, “occupied”, and “unknown” states (Bauer et al., 2020). Geometric rules such as cone models for radar specify explicit assignments of these masses according to the detection’s location and beam parameters.

In structured regression or uncertainty quantification, the geometric evidential head outputs the parameters of a Normal–Inverse–Gamma (NIG) distribution,

• $\mu_i$ — predicted mean
• $\lambda_i > 0$ — mean strength
• $\alpha_i > 1$ — variance shape
• $\beta_i > 0$ — variance scale forming the joint density

$$p(y,\sigma^2 \mid \mu,\lambda,\alpha,\beta) = \Bigl(\frac{\lambda}{2\pi\sigma^2}\Bigr)^{1/2} \exp\Bigl[-\frac{\lambda(y-\mu)^2}{2\,\sigma^2}\Bigr] \cdot \frac{\beta^\alpha}{\Gamma(\alpha)} (\sigma^2)^{-(\alpha+1)} \exp[-\beta/\sigma^2]$$

with the predictive variance decomposed into aleatoric $\beta/(\alpha-1)$ and epistemic $\beta/(\lambda(\alpha-1))$ components (Shihab et al., 12 Jan 2026).

2. Geometric Evidence Assignment in Sensor Models

For analytical sensor models such as radar cones:

• Cells prior to the detection along the beam ($|\theta_{ij}-\theta_k|\le \Delta\theta/2$ and $r_{ij} < r_k - \delta_r$): $m_{f,ij}^{(k)} = \lambda_f$, $m_{o,ij}^{(k)} = 0$, $m_{u,ij}^{(k)} = 1-\lambda_f$
• Cell at detection ($|\theta_{ij}-\theta_k|\le \Delta\theta/2$ and $|r_{ij}-r_k| \le \delta_r$): $m_{o,ij}^{(k)} = \lambda_o$, $m_{f,ij}^{(k)} = 0$, $m_{u,ij}^{(k)} = 1-\lambda_o$
• Cells outside the beam or beyond range: $m_{ij}^{(k)} = [0,0,1]^T$

Typical parameters for radar: $\lambda_f = 0.3$, $\lambda_o = 0.5$, with cones cast over $N=10$ sweeps and mass values assigned according to hand-crafted spatial logic (Bauer et al., 2020).

In domain-informed evidence heads (CalPro for protein structure), node and edge features reflecting geometric configuration (e.g., 3D distances, local structure, disorder priors) are propagated via graph neural network (GNN) message-passing, with output mappings enforcing $\alpha_i>1$ (via softplus), and encoding geometric priors as features that modulate the epistemic variance in the local NIG posterior (Shihab et al., 12 Jan 2026).

3. Evidential Fusion Rules and Update Mechanisms

Evidence masses from independent sources are combined according to formal rules:

• Dempster’s Rule (for geometric heads): combines mass vectors $m^{(1)}$, $m^{(2)}$ with conflict $K = m_{f1}\,m_{o2} + m_{o1}\,m_{f2}$,

$$m^{(1)} \oplus_D m^{(2)} = \frac{1}{1-K}\begin{bmatrix} m_{f1} m_{f2} + m_{f1} m_{u2} + m_{u1} m_{f2} \ m_{o1} m_{o2} + m_{o1} m_{u2} + m_{u1} m_{o2} \ m_{u1} m_{u2} \end{bmatrix}$$

ensuring monotonic reduction of unknown mass and convergence to certainty (Bauer et al., 2020).

• Yager’s Rule (for fusion with deep ISM masses): stores conflict in the unknown bin,

$$m^{(1)} \oplus_Y m^{(2)} = \begin{bmatrix} m_{f1} m_{f2} + m_{f1} m_{u2} + m_{u1} m_{f2} \ m_{o1} m_{o2} + m_{o1} m_{u2} + m_{u1} m_{o2} \ m_{u1} m_{u2} + K \end{bmatrix}$$

resolving clashes between prior and observation conservatively.

These rules are iteratively applied across time steps or layers, with map updates at each cell given by

$m_{ij}^{(k)} = m_{ij}^{(k-1)} \oplus m_{ij}^{(k)}$

using the appropriate rule depending on evidence source.

4. Discount Factors and Certainty Bounds

When blending geometric heads with learned priors or deep ISMs, influence is controlled by:

• Discount factor $\gamma$: regulates the contribution of the deep ISM evidence, applied as

$$\gamma \otimes \tilde{m} = \begin{bmatrix} \gamma\,\tilde{m}_f \ \gamma\,\tilde{m}_o \ 1-\gamma + \gamma\,\tilde{m}_u \end{bmatrix}$$

• Lower-bound guarantee: $$\gamma^{(k)} = \min\left\{ \tanh\big(\alpha \max(0, \Delta m_u)\big),\, \frac{m_u^{(k-1)} + K - \underline{m}_u}{m_u^{(k-1)}(1-\tilde{m}_u^{(k)})} \right\}$$ ensures $m_u \ge \underline{m}_u$ for all steps, preventing deep priors from overpowering evidence already allocated by geometric measurements. This mechanism enforces certifiable reliability of geometric evidence in contested or ambiguous regions (Bauer et al., 2020).

5. Convergence Analysis and Empirical Implications

Repeated application of geometric evidential head rules under Dempster’s fusion implies:

• Monotonicity: $$m_u^{(n)} \le m_u^{(n-1)} \implies \lim_{n\to\infty} m_u^{(n)} = 0$$ for cells observed by sensors, as free or occupied mass accumulates (Bauer et al., 2020).
• Sharp boundaries: On surface or hit cells, $m_o \to 1$; in free space, $m_f \to 1$, resulting in a step function boundary between states.
• Finite revisiting: Associativity and idempotence yield rapid convergence in revisited regions; map stabilization is guaranteed under sufficient sensor coverage.
• Conflict handling: Yager’s rule ensures that clashes with deep priors increase unknown mass, which is subsequently resolved in favor of geometric evidence without regressing certainty below $\underline{m}_u$.

Empirical ablations for protein uncertainty quantification validate the necessity of the evidential head. Removal increases sharpness intervals and worsens calibration error (ECE from 0.021 to 0.025; sharpness from 1.42Å to 1.98Å), confirming that the geometric evidential component is critical for tight, calibrated uncertainty (Shihab et al., 12 Jan 2026). This suggests a direct influence of the geometric evidential head on the reliability and precision of predictive intervals in spatially structured tasks.

6. Influence of Geometric Features in Graph-Based Architectures

Geometric evidential heads in GNN architectures, exemplified in protein structure modeling, integrate spatial connectivity, geometric edge features (e.g., Cα–Cα distances), and structural node annotations (secondary structure, disorder indices) into the evidence aggregation process. These features modulate local uncertainty through direct impact on outputted NIG parameters—high disorder or flexibility drives elevated epistemic variance, while well-packed regions yield sharper, lower-uncertainty outputs. The use of geometric priors as node attributes allows context-aware regularization and improved robustness to distribution shift or missing data (Shihab et al., 12 Jan 2026).

7. Synergy between Geometric and Deep Evidential Components

Combining geometric evidential heads with deep data-driven priors achieves both rapid initialization and ultimate precision in map boundaries or uncertainty intervals. The geometric component guarantees sharp separation and monotonic convergence under valid measurements; the deep prior efficiently fills in unobserved regions with moderate certainty. Restricting deep prior influence via discounting and lower-bound constraints, followed by conservative fusion (Yager’s rule), ensures that true evidence cannot be overwritten by the prior, facilitating fast, robust, and certifiable scene coverage or uncertainty estimation. The synthesis yields systems that combine coverage and geometric precision in both occupancy mapping and structured regression domains (Bauer et al., 2020, Shihab et al., 12 Jan 2026).