🧬 biology

Robust volatility updates for Hierarchical Gaussian Filtering

This paper introduces a robust modification to the Hierarchical Gaussian Filtering (HGF) algorithm that prevents negative posterior precision errors in volatility-coupled nodes by interpolating between two quadratic expansions of the variational energy using the Lambert W function, thereby ensuring stable belief updates across the entire parameter space.

Original authors: Christoph Mathys, Nicolas Legrand, Peter Thestrup Waade, Nace Mikus, Lilian Aline Weber

Published 2026-05-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christoph Mathys, Nicolas Legrand, Peter Thestrup Waade, Nace Mikus, Lilian Aline Weber

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to predict the weather. You have a mental model of how the sky works. Usually, your model is pretty good: you see a cloud, you think "rain," and you're right. But sometimes, the world throws you a curveball. A massive storm appears out of nowhere, or the sky suddenly clears up in a way your model never expected.

In the world of computer science and psychology, Hierarchical Gaussian Filtering (HGF) is a sophisticated tool used to model how agents (like humans or AI) update their beliefs when they see new information. It's like a multi-layered weather station where one layer predicts the temperature, and the layer above it predicts how fast the temperature is changing (the volatility).

However, the original version of this tool had a critical flaw: it could crash.

The Problem: The "Impossible" Math

Think of the HGF as a calculator trying to figure out how confident it should be in its prediction. This confidence is called "precision" (the opposite of uncertainty).

High precision = "I am very sure."
Low precision = "I have no idea."

In the original HGF, when a surprise was too big (a huge prediction error), the math would sometimes spit out a negative confidence. In the real world, you can't have "negative confidence." It's like saying you are -50% sure. It's a logical impossibility. When the computer tried to calculate this, it would hit a wall, throw an error, and stop working. This meant the model couldn't handle extreme situations, limiting how much it could learn.

The Solution: A Smarter Way to Approximate

The authors of this paper, Christoph Mathys and his team, fixed this by changing how the computer "approximates" the answer.

Imagine you are trying to draw a perfect circle on a piece of paper, but you only have a ruler.

The Old Way: You tried to draw a straight line that touches the circle at one point. If the circle was very bumpy or weirdly shaped (due to a huge surprise), your straight line would end up pointing in the wrong direction, leading to that "negative confidence" crash.
The New Way (uHGF): The authors realized that instead of just looking at the immediate point, they needed to look at the shape of the curve in two different ways and then blend them together.

They used a clever mathematical trick involving something called the Lambert W function. Think of this as a special "magic key" that allows the computer to instantly find a second, hidden peak in the data without having to search for it slowly.

Here is how their new method works in three simple steps:

Step 1: The Safe Bet. First, they calculate a "safe" estimate using only the parts of the math that are guaranteed to be stable. This ensures the confidence score never goes negative. It's like having a safety net.
Step 2: The Hidden Peak. If the data is really weird (like a massive storm), the math might actually have two possible answers (two peaks). The old method missed the second one. The new method uses the "magic key" (Lambert W) to instantly find where this second peak is.
Step 3: The Blend. Finally, they don't just pick one or the other. They mix the "Safe Bet" and the "Hidden Peak" together, like mixing two colors of paint, to create a final answer that is accurate, stable, and never crashes.

Why This Matters

The paper demonstrates that this new method, which they call uHGF (unbounded HGF), is much more robust.

It doesn't crash: In tests, the old method failed about 6% of the time when things got extreme, while the new method succeeded 100% of the time.
It handles the extremes: When the "weather" changes violently, the new model keeps working, adjusting its beliefs smoothly instead of breaking.
It opens up new possibilities: Because the model no longer has "forbidden zones" where it crashes, researchers can now explore a much wider range of scenarios and parameters without worrying about the computer stopping.

The Bottom Line

The authors didn't invent a new type of weather; they just fixed the weather station so it doesn't break when a tornado hits. By using a smarter mathematical approximation and a "magic key" to find hidden patterns, they made the Hierarchical Gaussian Filter robust enough to handle the wildest surprises the data can throw at it, ensuring the learning process never stops.

Technical Summary: Robust Volatility Updates for Hierarchical Gaussian Filtering

Problem Statement

Hierarchical Gaussian Filtering (HGF) is a variational inference framework used to model how agents update beliefs about hidden states in their environment. In HGF networks, parent nodes can influence the mean or the variance (volatility) of child nodes. While the update equations for value coupling (mean-targeting) are stable, the original update equations for volatility coupling (variance-targeting) suffer from a critical flaw: under specific conditions in the parameter space, they can yield negative posterior precision (inverse variance).

Since the variance of a Gaussian distribution must be positive, negative precision is a logical impossibility. In practice, this causes the filtering algorithm to terminate with an error. This issue arises when the quadratic approximation of the variational energy—used to derive the Gaussian posterior—becomes inappropriate. Specifically, when a large prediction error occurs relative to the previous uncertainty (a high ratio of total posterior uncertainty $\beta$ to previous variance $\alpha$ ), the variational energy develops a convex region or a second local maximum. Expanding the energy quadratically at the prior prediction point in these regions leads to a parabola opening upward, resulting in negative precision. This imposes artificial boundaries on the usable parameter space, forcing parameter estimation algorithms to avoid valid regions or crash during optimization.

Methodology

The authors propose a modified approach to the variational energy approximation that ensures robustness across the entire parameter space. The methodology involves three key steps:

Decomposition and Concave Approximation ( $L_1$ ):
The variational energy $J$ is decomposed into three components: $J_1$ (log-normalizer), $J_2$ (scaled uncertainty), and $J_3$ (quadratic prior). The authors identify that $J_2$ is responsible for the convex regions causing negative precision. To guarantee positive precision, they construct a first quadratic expansion ( $L_1$ ) based on a modified energy function $K = J_1 + J_3$ , effectively excluding the problematic $J_2$ term when calculating the curvature (precision). Since $K$ is strictly concave, the resulting precision $\pi_{L1}$ is always positive.
Closed-Form Second Mode Detection ( $L_2$ ):
In regimes where the variational energy becomes bimodal (due to large prediction errors), a single Gaussian centered at the prior is insufficient. The authors derive a closed-form approximation for the location of the second local maximum ( $x^*$ ) using the Lambert W function ( $W_0$ ). This approximation is exact in the limit where the previous uncertainty approaches zero ( $\alpha \to 0$ ), which is precisely the regime where bimodality arises. A second quadratic expansion ( $L_2$ ) is then constructed at this estimated mode $x^*$ , ensuring positive precision by utilizing the full curvature of the original energy function at that point.
Softmax-Weighted Mixture Moment Matching:
To produce a single Gaussian posterior that faithfully represents the potentially bimodal variational distribution, the two expansions ( $L_1$ and $L_2$ ) are combined. The authors assign weights to each expansion based on the variational energy evaluated at their respective means, using a softmax function. The final posterior mean and variance are computed via moment matching of this two-component Gaussian mixture. This ensures the resulting posterior is a valid Gaussian with strictly positive precision, regardless of the relative heights of the modes.

The framework is further extended to the Generalized HGF (gHGF), handling nodes with multiple volatility children and mixed value/volatility coupling, by applying the Lambert W strategy independently for each child and combining the resulting expansions.

Key Contributions

Robustness: The proposed update equations eliminate the possibility of negative posterior precision, allowing HGF networks to operate across the entire parameter space without crashing.
Analytical Solution: The use of the Lambert W function provides a computationally efficient, closed-form method for locating secondary modes, avoiding the need for iterative root-finding algorithms.
Faithful Approximation: The method preserves the core features of the original HGF (quadratic approximation, precision-weighted prediction errors) while significantly improving the accuracy of the Gaussian approximation in high-error regimes.
Generalizability: The approach is shown to extend naturally to the generalized HGF, supporting complex network topologies with multiple children and mixed coupling types.

Results

The authors validated the new equations (termed uHGF for unbounded HGF) through four comparative simulations:

Approximation Quality: Across 488 parameter combinations, the classic HGF failed (negative precision) in 6.4% of cases. The uHGF succeeded in 100% of cases. Where both succeeded, the uHGF reduced the Kullback–Leibler (KL) divergence between the approximate and true variational posterior by nearly two orders of magnitude (from 1.34 to 0.023).
Standard Conditions: Under standard filtering conditions with moderate prediction errors, both methods produced valid trajectories, with uHGF showing slightly improved tracking due to better approximation.
Extreme Prediction Errors: In simulations with high meta-volatility causing large prediction errors, the classic HGF crashed at step 115 due to negative precision. The uHGF successfully completed all 320 steps, maintaining strictly positive precision and adapting to regime changes.
Parameter Space Coverage: A systematic scan of 32,761 parameter combinations revealed that the classic HGF failed in 22.7% of cases (forming a large connected region of failure). The uHGF succeeded in 100% of cases, effectively removing the boundaries that previously restricted parameter estimation.

Significance

The paper claims that these modifications make HGF networks robust across the entire expanse of parameter space. The primary significance lies in the facilitation of parameter estimation. In computational psychiatry and cognitive neuroscience, fitting HGF models to behavioral data often involves optimization algorithms that explore the parameter space. The original negative precision issue created "hard boundaries" where the likelihood function was undefined, causing optimizers to fail or converge to boundary artifacts. By ensuring that every parameter set yields a valid filtering trajectory and a well-defined likelihood, the new equations allow for more reliable, faster, and unconstrained parameter estimation. This enables the use of a wider range of models and removes the need for costly boundary checks during inference.