Unified Uncertainty Propagation

• Unified uncertainty propagation is a framework that integrates probabilistic and statistical principles to track and quantify uncertainty across sequential modeling stages.
• It consistently represents all sources of uncertainty—measurement noise, model approximation, and process error—ensuring calibrated outputs and robust risk assessment.
• Practical implementations employ uncertainty-aware training, synthetic noise injection, and Jacobian-based error propagation to improve accuracy and calibration in complex systems.

Unified uncertainty propagation refers to a principled mathematical and algorithmic framework for tracking, transforming, and quantifying uncertainty as it flows through multi-stage ML, scientific modeling, or physical simulation chains. The guiding objective is to ensure that all relevant sources of uncertainty—measurement noise, model approximation, process error—are consistently represented, propagated, and calibrated through sequential or coupled computational steps, enabling reliable confidence and risk assessment in downstream inferences. Unified approaches supersede ad hoc or stepwise uncertainty treatments by embedding distributional information explicitly at each stage and by grounding propagation rules in rigorous probabilistic, statistical, or dynamical principles.

1. Probabilistic Formulation of Chained Model Uncertainty

Consider a chain of $k$ sequential models $f_1,\dots,f_k$ with initial random input $X\sim p_X(x)$. Each stage outputs $z_j = f_j(z_{j-1})$ for $j=1,\dots,k$ (define $z_0=X$). In a fully probabilistic setting, each map is promoted to a Markov kernel $p_j(z_j|z_{j-1})$. The joint law is: $$p_{Z_0,\ldots,Z_k}(z_0,\ldots,z_k) = p_X(z_0)\prod_{j=1}^k p_j(z_j|z_{j-1})$$ and the marginal predictive law for the final output variable $Y=z_k$ is: $$p_Y(y) = \int \cdots \int p_k(y|z_{k-1})\dots p_1(z_1|x)p_X(x)dx\,dz_1\cdots dz_{k-1}$$ For two-stage chains, the above reduces to a double integral: $f_1,\dots,f_k$0 This structure allows the explicit tracking of uncertainty transformation and accumulation through each model $f_1,\dots,f_k$1, capturing how errors or variability at stage $f_1,\dots,f_k$2 propagate to all downstream outputs (Douglas et al., 2024).

2. Uncertainty-Aware Model Construction and Training

To propagate input uncertainty rigorously, each model $f_1,\dots,f_k$3 in the chain must become "uncertainty-aware": it must accept a distributional input (typically parameterized by mean $f_1,\dots,f_k$4 and covariance $f_1,\dots,f_k$5 or per-feature standard deviations) and yield as output a predictive distribution, characterized analogously by $f_1,\dots,f_k$6. In neural networks, this is realized by:

• Appending input uncertainty (e.g., featurewise noise scale $f_1,\dots,f_k$7) to the network inputs.
• Configuring the network head to output both predictive mean and covariance.
• Training via likelihood maximization: e.g., for outputs $f_1,\dots,f_k$8, minimizing $f_1,\dots,f_k$9 if $X\sim p_X(x)$0 is Gaussian.
• Covariance transformation between stages via the chain of Jacobians: if $X\sim p_X(x)$1, $X\sim p_X(x)$2, then

$X\sim p_X(x)$3

This provides a classical error-propagation mechanism throughout the chain (Douglas et al., 2024).

3. Synthetic Noise Injection and Practical Propagation

Unified frameworks address the lack of ground-truth upstream uncertainty by synthetically injecting calibrated noise:

• For each feature $X\sim p_X(x)$4, sample $X\sim p_X(x)$5, set $X\sim p_X(x)$6, and perturb $X\sim p_X(x)$7 by $X\sim p_X(x)$8.
• Train "blind" models with only $X\sim p_X(x)$9 and "uncertainty-aware" models with $z_j = f_j(z_{j-1})$0.
• This allows empirical assessment of the benefit from input-uncertainty information, as shown by significant accuracy/loss/entropy improvements in edge and node GNN tasks when input uncertainty is provided and utilized proportionally to its magnitude (accuracy improvement from 0.846 to 0.864; binary cross-entropy decreased from 0.333 to 0.301 for $z_j = f_j(z_{j-1})$1 noise) (Douglas et al., 2024).

4. Extension to Multi-Stage and High-Dimensional Chains

For deeper chains, the principles generalize as follows:

• Each module implements $z_j = f_j(z_{j-1})$2, is trained to propagate both mean and covariance, and ingests uncertainty from upstream.
• When exact integration is infeasible, Gaussian (moment-based) or low-order polynomial approximations (e.g., via Taylor expansion or moment-matching) are maintained throughout.
• In high-dimensional representations, diagonal or low-rank plus diagonal covariance structures are recommended for tractability.
• For non-differentiable components, sampling-based propagation (unscented transform, Monte Carlo) or algebraic moment matching provides alternatives (Douglas et al., 2024).

5. Evaluation Metrics, Calibration, and Empirical Results

Unified propagation requires multi-metric evaluation, generally along:

• Prediction accuracy (fraction correct).
• Negative log-likelihood (calibrated loss).
• Mis-calibration area (MCA), measuring the area between confidence and empirical accuracy curves.
• Predictive entropy (sharpness). Experiments indicate that unified propagation not only improves point prediction and predictive confidence, but does so without degrading calibration, and the effect is noise-dependent: improvements only emerge above a critical input noise threshold ($z_j = f_j(z_{j-1})$3 in synthetic settings) (Douglas et al., 2024).

Empirical Results Table

Metric	Edge task Blind	Edge task UA	Node task Blind	Node task UA
Accuracy	0.846	0.864	0.923	0.928
Loss	0.333	0.301	—	—
MCA	0.0135	0.0161	—	—
Entropy	0.482	0.437	0.275	0.256

6. Best Practices and Implementation Considerations

• Every layer/module in the network or chain must become uncertainty-aware: accepting and outputting (mean, covariance) or (mean, variance) tuples.
• Calibration (e.g., via reliability diagrams or MCA), sharpness (output entropy), and accuracy should be monitored and maintained.
• If upstream uncertainty information is unavailable, synthetic noise injection during training enables downstream models to exploit uncertainty information when presented.
• In non-linear or non-differentiable systems, resort to sampling, unscented transform, or moment-matching to enable Gaussian or low-moment summaries to be propagated.
• For large feature spaces or high-dimensionality, diagonal or structured low-rank covariances are necessary for computational tractability.
• The Jacobian chain for error propagation is feasible only when the models are differentiable and not overly stiff; in other cases, downstream approximation or robust sampling should be considered (Douglas et al., 2024).

7. Significance and Broader Context

Unified uncertainty propagation frameworks are essential in reliability-critical settings—such as in high-energy physics event reconstruction—where sequential ML architectures are prevalent and confidence in each stage's output is imperative. By enforcing propagation of uncertainty in a distributional sense throughout the chain, one achieves:

• Transparency in how uncertainty is absorbed, transformed, and accumulated through all layers.
• Robust, calibrated, and sharper predictions, enabling more reliable downstream decision-making.
• Potential improvements in raw accuracy, particularly when input and upstream uncertainties are non-negligible and signal-rich. These principles apply broadly to chained ML systems, scientific pipelines, and any sequential inference architecture requiring rigorous uncertainty accounting and performance guarantees (Douglas et al., 2024).