📊 statistics

Recursive Entropic Variational Inference for Nonlinear State-Space Models

This paper introduces a class of recursive forward-backward algorithms for state estimation in nonlinear, non-Gaussian state-space models, derived from a variational Lagrangian formulation that utilizes entropic trust-region updates, Gauss-Markov approximations, and generalized statistical linear regression to achieve favorable computational complexity.

Original authors: Hany Abdulsamad, Ángel F. García-Fernández, Simo Särkkä

Published 2026-07-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hany Abdulsamad, Ángel F. García-Fernández, Simo Särkkä

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Navigating a Foggy Journey

Imagine you are trying to track a lost hiker in a dense, foggy forest. You have a map (your prior knowledge about where they might be) and you occasionally hear a shout or see a flash of a jacket (your noisy observations). However, the forest is tricky: the terrain twists and turns unpredictably (nonlinear), and the weather changes in ways you can't perfectly predict (non-Gaussian).

Your goal is to figure out exactly where the hiker was at every single moment in time, not just where they are right now. This is called smoothing.

The problem is that the math required to calculate the perfect answer is often impossible to solve directly. It's like trying to solve a maze while the walls are moving. So, scientists use "approximations"—best guesses that are good enough to be useful.

This paper proposes a new, smarter way to make those guesses.

The Core Idea: "Trust-Region" Hiking

The authors treat the problem of finding the hiker's path as a dynamic optimization problem. Instead of jumping straight to a final answer, they imagine taking a series of small, careful steps.

Here is the analogy they use: The Trust-Region Hike.

The Current Guess: You start with a rough idea of where the hiker is (your current "posterior" distribution).
The Constraint (The Rope): You are tied to your current guess by a rope of a specific length. This is the KL-divergence constraint (or "trust region"). It means you are allowed to change your guess, but only so far. You can't jump wildly to a completely different location; you must stay within a "safe zone" around your current belief.
The Entropic Pull: Inside this safe zone, you want to move toward the true answer (the hiker's actual path). The paper uses a mathematical concept called entropy to measure how much "information" or "surprise" is involved in moving from your old guess to the new one.
The Damping Parameter (The Step Size): This is the most clever part. The paper introduces a "damping" knob (called $\beta$ $β$ ).
- If the rope is tight (high damping), you take tiny, cautious steps.
- If the rope is loose (low damping), you can take bigger leaps.
- The algorithm automatically figures out the perfect step size for each move to ensure you don't overshoot the target or get stuck in a loop.

The Three Ways to Walk the Path

The paper shows that you can organize this "hike" in three different ways, depending on how you look at the timeline. They call these Factorizations:

Forward-Markov (The "Looking Back" Hike):
- Imagine you start at the beginning of the hike and move forward in time.
- You first look backward from the end of the trail to gather all the clues about where the hiker might have come from.
- Then, you walk forward again, updating your map step-by-step using those backward clues.
- Analogy: Like reading a mystery novel from the last page to the first to understand the plot, then reading it again normally with that new knowledge.
Reverse-Markov (The "Looking Forward" Hike):
- This is the opposite. You start at the end of the trail and work your way backward to the start.
- You gather clues about the future first, then walk backward to update your understanding of the past.
- Analogy: Like a detective starting at the crime scene and rewinding time to figure out how the suspect got there.
Hybrid (The "Two-Way" Hike):
- This combines both. You send a scout forward and a scout backward simultaneously.
- They meet in the middle, sharing their clues. The final map is a blend of what the forward scout saw and what the backward scout saw.
- Analogy: Like two people walking toward each other on a bridge, meeting in the middle to compare notes on the view from both sides.

Handling the "Messy" Forest (Nonlinearity)

In a perfect world, the forest would be flat and straight (Linear-Gaussian). In that case, the math is easy, and you get the exact answer (like the famous Rauch–Tung–Striebel smoother).

But real forests are messy. The trees are crooked, and the wind blows sideways.

The Problem: When the forest is messy, the math gets too hard to solve exactly.
The Solution: The paper uses two "tricks" to approximate the messy forest with a simpler, straighter version just for the current step:
1. Generalized Statistical Linear Regression (GSLR): This is like drawing a straight line through a curve. It approximates the complex, twisting path with a straight line that fits the current data points.
2. Fourier–Hermite Expansion: This is a more sophisticated trick. Instead of just drawing a straight line, it uses a series of curved shapes (like waves) to fit the twisty path much more accurately. The paper shows this second method is better at handling extreme twists and turns.

Why This Matters

The authors claim their method is principled. Many existing methods are "ad hoc," meaning they are patched together with rules of thumb that work sometimes but aren't mathematically guaranteed to be stable.

Their method is built on a solid mathematical foundation (Variational Inference and Proximal Optimization). Because of this:

It's Stable: The "rope" (trust region) prevents the algorithm from going crazy and making huge, wrong jumps.
It's Efficient: Even though it's complex, the math is structured so it can be solved quickly, step-by-step, without needing to re-calculate the whole history every time.
It's Flexible: It works for simple straight paths and for the most twisted, non-Gaussian forests.

The Results (What the Paper Found)

The authors tested their "hiking" algorithms on three scenarios:

A Perfect Forest (Linear-Gaussian): They showed that if the forest is simple, their method finds the exact same answer as the classic, perfect method. This proves their math is correct.
A Volatile Forest (Stochastic Volatility): They simulated a scenario where the "wind" (volatility) changes wildly. They found that the "straight line" approximation (GSLR) failed to track the hiker because it couldn't see the wind's effect. However, the "curved shape" approximation (Fourier–Hermite) tracked the hiker perfectly.
A Cubic Sensor (Cubic Sensor Problem): They tested a scenario where the clues are distorted by a cube (a very weird, non-linear shape). Their method converged smoothly to the right answer, while older, "undamped" methods got stuck in a loop, oscillating back and forth without ever settling.

Summary

In short, this paper gives us a new, robust toolkit for tracking things in a chaotic, unpredictable world. Instead of guessing wildly, it takes small, mathematically guaranteed steps, constantly checking its own confidence to ensure it doesn't lose its way. It offers three different "routes" (Forward, Reverse, Hybrid) to solve the same problem, ensuring that no matter how the data is structured, there is a reliable way to find the truth.

Technical Summary: Recursive Entropic Variational Inference for Nonlinear State-Space Models

Problem Statement

The paper addresses the challenge of accurate state estimation in nonlinear, non-Gaussian state-space models (SSMs). While Bayesian inference provides a principled framework for such problems, exact recursive solutions (like the Rauch–Tung–Striebel smoother) are generally unavailable for non-conjugate, nonlinear systems. Existing approximate methods, such as sequential Monte Carlo, suffer from high computational complexity and poor scalability in high-dimensional spaces. Conversely, many existing variational inference (VI) and expectation propagation (EP) approaches for SSMs extend these frameworks in an ad hoc manner, often embedding variational principles into the Rauch–Tung–Striebel template rather than deriving them from a rigorous dynamic optimization problem. This limits their ability to fully leverage systematic approaches to state estimation.

Methodology

The authors propose a unified framework that casts approximate Bayesian inference in SSMs as an iterative dynamic optimization problem over the space of candidate posterior distributions. The core methodology relies on Entropic Proximal Variational Optimization.

1. Variational Lagrangian Formulation

The inference problem is formulated as a sequence of entropic trust-region updates. Instead of directly maximizing the Evidence Lower Bound (ELBO), the method solves a constrained optimization problem where each update is subject to a Kullback–Leibler (KL) divergence constraint relative to the previous iterate. This introduces a "damping" mechanism controlled by a Lagrange multiplier (or step size parameter $\beta$ ).

Dynamic Optimization: The problem is decomposed using forward-Markov and reverse-Markov factorizations of the variational posterior.
Tilted Distributions: The solution to the constrained optimization yields "tilted" distributions, which are geometric interpolations between the exact posterior and the previous iterate. The interpolation parameter $\beta$ is determined by minimizing a dual objective related to the Rényi divergence.

2. Recursive Algorithms via Gauss–Markov Approximations

To render the abstract recursions tractable, the authors restrict the variational family to Gauss–Markov approximations. This leads to three distinct recursive algorithms:

Forward Entropic Variational Smoother (FEVS): Utilizes a forward-Markov factorization, performing a backward recursion to accumulate future measurement potentials followed by a forward propagation of marginals.
Reverse Entropic Variational Smoother (REVS): Utilizes a reverse-Markov factorization, performing a forward recursion for past-measurement potentials followed by backward propagation.
Hybrid Entropic Variational Smoother (HEVS): Combines both directional recursions, allowing intermediate marginals to receive information from both temporal directions simultaneously.

3. Practical Realization

Since the exact potentials in nonlinear models are intractable, the framework employs two strategies to approximate the log-densities of the dynamics and measurement models:

Generalized Statistical Linear Regression (GSLR): Constructs affine-Gaussian surrogate models by matching conditional moments.
Fourier–Hermite Moment Matching: Uses second-order Fourier–Hermite series expansions to capture higher-order curvature information without assuming additive Gaussian noise.

These approximations allow the potential functions and log-normalizers to be represented as quadratic forms, enabling closed-form recursive updates for the mean and covariance parameters.

4. Adaptive Step Size

A critical component of the method is the adaptive selection of the damping parameter $\beta$ . The paper proposes using a bisection method on the Lagrange multiplier to satisfy the KL trust-region constraint ( $\varepsilon$ ). This ensures that updates remain within a region where local approximations are valid, providing a principled mechanism for step-size control that prevents divergence in highly nonlinear regimes.

Key Contributions

Dynamic Optimization Formulation: The paper introduces a principled variational framework that formulates inference as a sequence of entropic trust-region updates, moving beyond ad hoc extensions of existing smoothers.
Unified Recursive Algorithms: It derives a family of forward-backward recursion schemes (FEVS, REVS, HEVS) whose structure is entirely determined by the factorization of the variational posterior.
Adaptive Step Size: The authors demonstrate that the step size corresponds to a Lagrange multiplier determined by a dual Rényi divergence objective, enabling automatic trust-region enforcement.
Generalization of Classical Smoothers: The approach generalizes classical linear-Gaussian smoothers (like RTS) and connects them to modern approximate inference techniques.
Flexible Approximations: The framework supports practical realizations using GSLR and Fourier–Hermite expansions, facilitating efficient inference in nonlinear and non-Gaussian models.

Results

The paper provides numerical evaluations on three benchmarks:

Linear-Gaussian Models: The algorithms are shown to converge to the exact Rauch–Tung–Striebel (RTS) solution, validating the correctness of the recursions. The damping parameter $\beta$ controls the convergence speed, with larger trust-region radii allowing faster convergence.
Stochastic Volatility (Nonlinear/Non-Gaussian): In a model where the observation depends on the state's variance, the Fourier–Hermite (FH) variant significantly outperforms the Generalized Statistical Linear Regression (GSLR) variant. The GSLR approach fails to capture the information in the variance (collapsing to the prior), whereas the FH variant successfully tracks the latent log-volatility.
Cubic Sensor Problem: In a scenario with a cubic measurement nonlinearity, the proposed entropic smoothers (HEVS) exhibit stable convergence of the Evidence Lower Bound (ELBO). In contrast, an undamped iterated posterior linearization smoother (IPLS) baseline fails to converge, oscillating between solutions.

Significance and Claims

The paper claims to provide a principled bridge between variational inference, proximal optimization, and recursive Bayesian inference. By framing inference as a dynamic optimization problem with entropic regularization, the method offers:

Stability: The KL-based trust-region constraint regularizes updates, reducing sensitivity to local approximation errors and preventing divergence in nonlinear settings.
Scalability: The resulting algorithms possess linear time complexity with respect to the time horizon $T$ , making them suitable for long sequences.
Theoretical Grounding: Unlike heuristic extensions of the RTS smoother, the proposed recursions arise as optimality conditions of a well-posed variational problem, offering a clearer path for convergence analysis and extension to richer variational families.

The authors position this work as a generalization of existing posterior-linearization and expectation propagation methods, resolving issues related to step-size control and the handling of non-additive noise through the introduction of entropic proximal constraints and higher-order moment matching.