⚡ electrical engineering

Joint Identification of Linear Dynamics and Noise Covariance via Distributional Estimation

This paper proposes a novel framework for the joint identification of linear system dynamics and noise covariance under general non-Gaussian distributions by leveraging state-transition distribution shape information through maximum likelihood and score-matching estimators, which demonstrate superior performance over ordinary least squares baselines.

Original authors: Yang Hu, Na Li

Published 2026-04-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Hu, Na Li

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

Imagine you are trying to teach a robot how to walk. You watch it take steps, recording where it is now ( $x$ ) and where it ends up next ( $x'$ ).

In the old days, scientists used a very simple method to figure out the robot's walking rules. They assumed the robot's wobbles (noise) were perfectly random, like a gentle, predictable breeze. They used a standard math tool called Ordinary Least Squares (OLS) to draw a straight line through the data.

The Problem:
Real life isn't a gentle breeze. Sometimes the robot slips on a banana peel (a huge outlier), or gets hit by a sudden gust of wind. These "wobbles" have a specific shape.

Gaussian (Old Way): Like a bell curve. Most wobbles are small, and huge ones are almost impossible.
Real World: Sometimes the wobbles are "heavy-tailed." This means small wobbles are common, but massive slips happen more often than the old math predicts.

If you use the old "bell curve" math on data with "heavy tails," your robot will learn the wrong rules. It will be too scared of the rare big slips and won't learn the true pattern. Also, the old method usually ignored how the noise was shaped; it just guessed the average.

The New Solution:
This paper proposes a new way to look at the data. Instead of assuming the noise is a simple bell curve, the authors say: "Let's assume the noise has a specific shape (like a Student-t distribution) and learn both the robot's rules AND the shape of the noise at the same time."

They call this Joint Identification.

The Creative Analogy: The Chef and the Soup

Imagine you are a chef trying to recreate a famous soup.

The System Dynamics ( $A$ ): This is the recipe (how much salt, how much water).
The Noise ( $\Sigma$ ): This is the "imperfection" in the kitchen. Maybe the stove flickers, or the chef's hand shakes.

The Old Way (OLS):
The chef tastes the soup and says, "Okay, the average taste was salty. I'll just add salt to match the average." They ignore the fact that sometimes the soup is way too salty because the stove flared up. They assume the "hand shake" is always the same gentle tremor.

The New Way (MLE & SME):
The new chefs (the authors) say: "Wait! The stove flares up occasionally, and the hand shake gets violent when we get tired. We need to figure out the recipe AND the specific pattern of the chaos."

They use two new tools:

MLE (Maximum Likelihood Estimator): Think of this as a detective who looks at the entire history of the soup. "If the stove flares up 10% of the time, and the soup tastes like this 90% of the time, what is the most likely recipe?" It uses the "shape" of the chaos to find the truth.
SME (Score-Matching Estimator): This is like a sculptor. Instead of looking at the final soup, they look at how the flavor changes as you add ingredients. They match the "slope" of the flavor curve to find the perfect recipe and the perfect description of the chaos.

Why is this a big deal?

It's Smarter: By understanding the "shape" of the noise (e.g., "Oh, big mistakes happen more often than we thought"), these new methods ignore the outliers that would trick the old method.
It's Faster to Learn: Because they use more information (the shape), they need fewer samples of data to get it right. It's like learning to drive: if you know exactly how slippery the road gets in the rain (the noise shape), you learn to drive safely in the rain much faster than if you just guessed.
It's Robust: Even if you guess the "shape" of the noise slightly wrong, these methods still work better than the old way.

The Trade-off (The Catch)

There is a price to pay for this intelligence.

The Old Way (OLS): Is like using a calculator. You punch in the numbers, and beep, you get the answer instantly. It's fast but dumb.
The New Way (MLE/SME): Is like solving a complex puzzle. You have to crunch numbers, run simulations, and iterate. It takes more computer power and time.

The Conclusion:
The paper shows that if you have a lot of data, the extra time is worth it because the new methods are much more accurate. If you are in a high-stakes situation (like controlling a self-driving car or a robot in a disaster zone), you don't want the "dumb" method that gets confused by a banana peel. You want the "smart" method that understands the shape of the chaos.

In short: The authors found a way to stop ignoring the "weird" parts of the data and use them to build a much better model of how the world works.

1. Problem Statement

The paper addresses the joint identification problem for linear time-invariant (LTI) systems. The goal is to simultaneously estimate two unknown parameters using state transition data $D = \{(x_i, x'_i)\}_{i=1}^N$ :

The Dynamical Matrix ( $A$ ): Governs the system's evolution.
The Noise Covariance Matrix ( $\Sigma$ ): Governs the scale of the stochastic noise.

System Model:
The system is defined as $x' = A x + \Sigma^{1/2} w$ , where $w$ is an i.i.d. unit noise with zero mean and identity covariance.

Key Constraint: Unlike classical approaches that assume Gaussian or sub-Gaussian noise, this work considers general noise distributions. The noise is modeled as a scaled version of a known "base density" $\phi(\cdot)$ .
The Gap: Existing methods typically handle dynamics identification (estimating $A$ ) and covariance identification (estimating $\Sigma$ ) separately, often relying on a two-stage approach (estimate $A$ first, then $\Sigma$ ). The authors argue that this tandem approach implicitly assumes Gaussian noise and fails to utilize the "shape" information (e.g., tail behavior, higher-order moments) inherent in non-Gaussian distributions, leading to suboptimal performance.

2. Methodology

A. Reparameterization via $\phi$ -Density Families

The core theoretical innovation is a novel reparameterization of the state-transition distribution.

Instead of treating $A$ and $\Sigma$ as separate entities affecting mean and variance independently, the authors define the conditional density of the next state $x'$ given $x$ as:
$f_{A,\Sigma}(x'|x) = \det(\Sigma)^{-1/2} \phi\left(\Sigma^{-1/2}(x' - Ax)\right)$
This formulation places the transition distribution within a $\phi$ -density family, parameterized jointly by $(A, \Sigma)$ .
Elliptical Distributions: To ensure computational tractability, the paper focuses on elliptical base densities (where $\phi(w) = \psi(w^\top w)$ ). This class includes Gaussian and Student-t distributions. For these families, the gradients and Hessians of the log-density can be expressed in closed forms involving scalar functions $\rho(z)$ and $\tau(z)$ .

B. Proposed Estimators

The authors propose two estimators that leverage the specific shape of the base density $\phi$ , contrasting them with the naive Ordinary Least Squares (OLS) estimator.

Maximum Likelihood Estimator (MLE):
- Maximizes the log-likelihood of the observed transitions under the $\phi$ -density family.
- Characterization: For elliptical families, the MLE solution resembles a reweighted OLS estimator. The weights depend on the residuals and the base density's shape function $\rho$ . This effectively down-weights outliers in heavy-tailed distributions (e.g., Student-t), improving robustness.
- Properties: Proven to be consistent (unbiased) and asymptotically normal.
Score-Matching Estimator (SME):
- Minimizes the squared difference between the score function (gradient of the log-density) of the estimated model and the true model.
- Utilizes an integration-by-parts trick to avoid needing the true density, making the loss function tractable.
- Characterization: Similar to MLE, the SME for elliptical families takes a reweighted OLS form but uses a different weighting scheme involving both $\rho$ and $\tau$ functions.

C. Theoretical Analysis

Consistency: Both MLE and SME are proven to uniquely converge to the ground-truth parameters $(A^*, \Sigma^*)$ as sample size $N \to \infty$ .
Sample Complexity: Under assumptions of system stability, sub-Gaussian noise, and bounded hypothesis spaces, the paper establishes a sample complexity of $\tilde{O}(d/N)$ for the dynamical matrix and $\tilde{O}(d/N)$ for the covariance matrix. This matches the optimal rate of standard MLEs.
Gaussian Limit: A key theoretical result (Lemma 9) shows that if the base density is Gaussian, both MLE and SME degenerate exactly to the naive OLS estimator. This confirms that the proposed methods only deviate from OLS when the noise distribution has non-Gaussian "shape" information to exploit.

3. Key Contributions

Novel Framework: First paper to formally study the joint identification of linear dynamics and noise covariance for generic (non-Gaussian) noise distributions, moving beyond the standard two-stage Gaussian assumption.
Reparameterization: Introduced a unified parameterization using $\phi$ -density families that encodes both dynamics and covariance into a single estimation problem, enabling the use of distributional shape information.
New Estimators: Proposed and analyzed two practical estimators (MLE and SME) that outperform OLS in non-Gaussian settings.
Theoretical Guarantees: Provided rigorous proofs for consistency, asymptotic normality, and finite-sample complexity bounds.
Iterative Algorithms: Developed efficient iterative algorithms for computing these estimators in the case of elliptical families, avoiding the need for complex non-convex optimization solvers in many cases.

4. Simulation Results

The authors validated their theory using simulations with Student-t distributed noise (a heavy-tailed distribution).

Performance vs. OLS:
- Accuracy: Both MLE and SME significantly outperformed the naive OLS baseline in estimating both $A$ and $\Sigma$ . MLE generally achieved the lowest error.
- Covariance Estimation: The improvement was most pronounced in estimating $\Sigma$ . OLS performed poorly because it failed to account for the heavy tails of the Student-t distribution, whereas MLE/SME correctly down-weighted outliers.
- Dimensionality: The performance gap between MLE and OLS widened as the system dimension increased, attributed to the increased dissimilarity between high-dimensional Student-t and Gaussian distributions.
Computational Trade-off:
- MLE and SME require iterative optimization, making them computationally more expensive than the closed-form OLS.
- However, the proposed iterative algorithm (reweighted OLS) converged quickly (within ~20 iterations), offering a practical trade-off between sample efficiency and computational cost.
Robustness: The estimators showed robustness to mild misspecification of the base density, though performance degraded if the assumed density was significantly different from the true noise distribution.

5. Significance and Impact

Beyond Gaussian Assumptions: This work challenges the standard practice in system identification of assuming Gaussian noise. It demonstrates that ignoring the "shape" of noise (e.g., heavy tails) leads to suboptimal identification, particularly for covariance estimation.
Downstream Applications: Accurate joint identification of dynamics and noise covariance is critical for risk-sensitive control, distributionally robust control, and adaptive Kalman filtering, where knowledge of the true noise structure is often assumed but rarely estimated correctly in non-Gaussian environments.
Theoretical Foundation: By establishing the connection between linear system identification and $\phi$ -density families, the paper opens a new avenue for applying statistical learning theory (like score matching) to control and system identification problems.

In summary, the paper provides a mathematically rigorous and empirically validated framework for identifying linear systems under general noise conditions, proving that leveraging distributional shape information yields superior estimation accuracy compared to traditional methods.