⚡ electrical engineering

Cyclic Reformulation-Based Identification and Polytopic Uncertainty Modeling for Multirate Systems

This paper proposes a non-iterative, control-oriented identification framework for multirate systems with missing outputs that utilizes cyclic reformulation to derive both a noise-reduced nominal model and a polytopic uncertainty model from a single dataset, thereby enabling robust LMI-based controller design without iterative optimization.

Original authors: Hiroshi Okajima, Kakeru Ono

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hiroshi Okajima, Kakeru Ono

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 to walk. To do this, you need to understand how its legs, joints, and sensors work together. But here's the twist: the robot has a mix of sensors. Some are super-fast cameras snapping pictures 60 times a second, while others are slower gyroscopes that only check in 20 times a second. Worse yet, sometimes the fast camera misses a frame entirely, leaving a gap in the data.

This is the messy reality of multirate systems. For a long time, figuring out the robot's true "brain" (its mathematical model) from this jumbled, missing data was a headache. Most existing methods tried to guess the missing pieces by drawing lines between the dots (interpolation) or by running endless, slow guess-and-check loops (iterative methods like EM algorithms). The authors of this paper say: "Nope, let's try something smarter and faster."

The Magic Trick: Unfolding Time

The team, led by Hiroshi Okajima and Kakeru Ono, uses a clever math trick called cyclic reformulation.

Imagine the robot's data as a long, tangled string of beads. Because the sensors work at different speeds, the beads are spaced out unevenly. Instead of trying to untangle them one by one, the authors' method takes the whole string and folds it into a neat, repeating pattern. They call this period M (which is just the smallest number of steps where all the sensors line up again).

By folding the data this way, they turn a messy, time-changing problem into a clean, steady one. They run a standard identification tool (called subspace identification) on this folded data. The result? Instead of getting one answer, the math spits out M different versions of the robot's model.

The "Group Project" Solution

Here is where the paper gets really clever. In a perfect, noise-free world, all M versions would be identical twins. But in the real world, noise (static, glitches, sensor errors) makes them look slightly different, like a group of friends trying to draw the same picture from memory.

The paper proposes two ways to use these M different drawings:

The Centroid (The "Average" Friend): The authors take all M drawings and average them out. Just like how averaging the guesses of a crowd often gets you closer to the truth than any single person, this "centroid" model cancels out the random noise. In their simulations, this average model was significantly better than trying to pick the single "best" drawing from the group. It also crushed the old "interpolation" methods, which were stuck around 67% accuracy no matter how much data they had, while the new average model hit over 96% accuracy even with noisy data.
The Polytope (The "Safety Bubble"): Instead of just picking one average, the authors use all M drawings to build a 3D shape (a polytope) that wraps around them. Think of it as a safety bubble. The true robot model is likely hiding somewhere inside this bubble. This is huge for engineers because it lets them design controllers that are guaranteed to work, even if the robot behaves like any of the M different versions inside the bubble.

What the Paper Rules Out

The authors are very clear about what their method is not.

It is not an iterative guess-and-check: They explicitly avoid the slow, computationally heavy methods that require you to guess a starting point and hope you don't get stuck in a local trap. Their method is non-iterative, meaning it runs once and gives you the answer.
It is not a "magic fix" for missing data: They don't claim to magically know what the missing sensor data was. Instead, they treat the missing data as zeros (a "zero-padded" sequence) and use the math to recover the system's structure without ever needing to fill in the blanks with fake numbers.
It is not a guaranteed "set-membership" bound: The paper is careful to say that the "safety bubble" (polytope) is built from the observed noise in the data, not from a theoretical guarantee that the true model is definitely inside. It's a data-driven description, not a mathematically proven cage.

How Sure Are They?

The authors are confident, but they stick to the evidence they have.

Simulations Only: All the results come from computer simulations. They tested a simple single-sensor system and a more complex multi-sensor system.
The Numbers: In their tests, even with high noise levels (standard deviation of 0.1), the "best" model found inside their safety bubble still predicted the robot's behavior with over 95% accuracy on average.
The "Best" Vertex: They found that the average model (centroid) consistently outperformed the single best individual model from the group. For example, in one test, the average hit 98.36% accuracy while the best single guess was only 95.55%.

The Bottom Line

This paper doesn't claim to have solved every problem in robotics. It doesn't work on physical robots yet (that's for future work), and it assumes the robot's basic physics don't change over time.

However, it offers a fresh, fast, and surprisingly accurate way to handle the messy reality of mixed-speed sensors. By folding time into a repeating pattern and then averaging the results, the authors show that you can get a super-clean model and a robust safety bubble without running endless loops of calculation. It's a "one-and-done" approach that turns a chaotic data stream into a reliable blueprint for control.

Technical Summary: Cyclic Reformulation-Based Identification and Polytopic Uncertainty Modeling for Multirate Systems

Problem Statement
Modern control systems frequently utilize heterogeneous sensors operating at different sampling rates (multirate systems). This creates a fundamental challenge for system identification: output measurements are intermittently missing, and the system behaves as a periodically time-varying (PTV) system rather than a standard linear time-invariant (LTI) system. Existing approaches face significant limitations:

Subspace and Lifting methods: While they can identify fast-rate models, recovering the original system parameters is difficult due to variable products in the expanded state space.
Iterative methods (e.g., EM algorithms): These treat missing outputs as latent variables but require careful initialization, suffer from local convergence issues, and are computationally intensive.
Interpolation-based methods: Filling missing data via interpolation introduces systematic modeling errors that limit accuracy regardless of noise levels.

The paper addresses the gap in providing a non-iterative, control-oriented identification method that yields both a robust nominal model and a polytopic uncertainty description from a single dataset, specifically handling the noise-induced discrepancies inherent in multirate identification.

Methodology
The proposed framework relies on cyclic reformulation, transforming the multirate system into an equivalent time-invariant representation with an expanded dimension. The methodology proceeds through the following steps:

System Representation: The multirate system is modeled as an LTI plant with an output selection matrix $V_k$ that varies periodically with period $M$ (the least common multiple of sensor sampling periods). Missing outputs are zero-padded for identification purposes.
Cyclic Reformulation: The periodic system is converted into a time-invariant system of dimension $Mn$ (where $n$ is the system order) using cyclic shift matrices. This allows the application of standard subspace identification methods (e.g., MOESP) to the "cycled" input-output signals.
Coordinate Transformation: Subspace identification yields dense matrices of dimension $Mn$. A specific coordinate transformation matrix $T$ $T$ is constructed to recover the cyclic reformulation structure. This step extracts $M$ $M$ distinct parameter sets $(A_{mi}, B_{mi}, C_{mi}, D_{mi})$ $(A_{mi}, B_{mi}, C_{mi}, D_{mi})$ for $i = 0, \dots, M-1$ $i = 0, \dots, M - 1$ .
- In a noise-free scenario, these $M$ sets coincide exactly (up to a coordinate transformation).
- Under finite noise, they exhibit dispersion due to noise-induced perturbations.
Dual-Model Construction: The $M$ $M$ parameter sets are utilized in two complementary ways:
- Centroid Model: The $M$ sets are averaged to form a single nominal model. For state matrices $A$ and $B$ , a uniform average is taken. For output matrices $C$ and $D$ , a row-wise average is computed using only the phases where the specific output was observed, avoiding double-counting of zero-padded data.
- Polytopic Uncertainty Model: The $M$ parameter sets serve as the vertices of a convex polytope. To ensure well-posedness, rows of $C$ and $D$ corresponding to non-observation phases are imputed with the row-wise centroid values. The resulting polytope is defined by the convex combination of these $M$ vertices.

Key Contributions

Non-Iterative Identification: The method provides an exact, non-iterative identification pipeline that recovers original system matrices (up to coordinate transformation) without the convergence risks of EM algorithms.
Exploitation of Noise Dispersion: Unlike previous work (e.g., [25]) which treated the $M$ parameter sets as coincident in the noise-free limit, this paper leverages their noise-induced dispersion to construct a polytopic uncertainty model. This model is directly compatible with vertex-based Linear Matrix Inequality (LMI) robust control design.
Noise Reduction via Centroid: The paper demonstrates that the centroid of the $M$ sets acts as a noise-reduced nominal model, outperforming individual vertices and interpolation-based baselines.
Structural Recovery: The paper restates the noise-free structural recovery theorem (Theorem 1) and separates it from the finite-noise modeling step, assessing the latter through output-level validation statistics rather than matrix-coordinate distances.

Results
Numerical simulations on SISO and MIMO multirate systems validate the approach:

Centroid Performance: In a SISO example with missing 2/3 of outputs, the centroid model achieved validation FIT scores above 96% across all noise levels, substantially outperforming a linear-interpolation-based MOESP baseline (which remained around 67%). The centroid consistently outperformed the best individual vertex, confirming the noise-reduction effect of averaging.
Polytope Reliability: In a MIMO example with two sensors ( $M=6$ ), the constructed polytope contained models with high predictive accuracy. Even at high noise levels ( $\sigma = 0.1$ ), the "best in-polytope" model (found via optimization) achieved average validation FITs above 95%.
Robustness: The polytope reliability remained high (>98%) across various sensor period configurations ( $M \in \{4, 6, 8, 12\}$ ) and data lengths, indicating the method's applicability to diverse multirate sensing scenarios.

Significance and Claims
The paper claims to bridge multirate system identification with robust control-oriented uncertainty modeling without iterative optimization. Its primary significance lies in:

Direct Compatibility: The resulting polytope is naturally compatible with vertex-based LMI robust control design (e.g., $H_\infty$ , robust pole placement), as the vertices are derived directly from the identification process.
Data-Driven Uncertainty: The polytope is presented as a "data-derived uncertainty description" rather than a guaranteed set-membership bound. The authors explicitly note that the vertices arise from the dispersion of estimates, and the polytope's utility is validated empirically through output-level statistics.
Modest Scope: The authors clarify that the theoretical guarantees (exact cyclic recovery) apply to the noise-free case. For finite noise, the method is an empirical approach where the polytope's effectiveness is demonstrated through numerical simulation rather than formal probabilistic bounds. The paper does not claim to solve systems with intrinsic parameter variations or nonlinearities, nor does it provide experimental verification on physical hardware, leaving these as future work.

In summary, the paper offers a practical, non-iterative pipeline that transforms the challenge of missing multirate data into an opportunity for generating both a high-fidelity nominal model and a structured uncertainty set suitable for robust control synthesis.