⚡ electrical engineering

Convergence Analysis of Natural Power Method and Its Applications to Control

This paper analyzes the convergence of the discrete-time natural power method to the dominant subspace defined by eigenvalues with the largest absolute values and leverages this property to develop model order reduction and low-rank controller synthesis techniques for both discrete-time LTI and slowly-varying LTV systems.

Original authors: Daiki Tsuzuki, Kentaro Ohki

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daiki Tsuzuki, Kentaro Ohki

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 standing in a crowded room full of people moving in different directions. Some are walking slowly, some are running, and some are dancing wildly. Your goal is to figure out the "main vibe" of the room—the most dominant patterns of movement—without getting overwhelmed by the noise of everyone else.

This paper is about a mathematical tool called the Natural Power Method (NPM) that helps computers do exactly that: find the most important "patterns" or "directions" in a complex system of data.

Here is a breakdown of what the authors discovered and how they used it, explained simply:

1. The "Biggest" vs. The "Most Real"

There is a famous, older method called the Oja Flow (think of it as a very popular, well-known GPS). This GPS is great at finding the directions where things are moving the most forward (mathematically, the largest "real" parts).

The authors introduce a new GPS: the Natural Power Method.

The Difference: While the old GPS looks for the biggest "forward" movement, the new NPM looks for the biggest total energy or magnitude (mathematically, the largest absolute values).
The Analogy: Imagine a spinning top. The Oja Flow might focus on how fast it's moving forward across the table. The Natural Power Method focuses on how violently it's spinning, regardless of which way it's drifting. For digital systems that work in steps (discrete-time), the "spin" (absolute value) is often the most critical factor to track.

2. How It Works (The "Squeeze")

The algorithm works like a magical sieve.

You start with a random guess of the main patterns.
The algorithm repeatedly "squeezes" the data through a filter.
With every squeeze, the noise (the weak, unimportant patterns) gets filtered out, and the strong, dominant patterns get stronger and clearer.
The Catch: The paper proves that this sieve works perfectly for any square grid of numbers, not just the special, "nice" grids that previous methods required. It converges to the $r$ most powerful patterns, provided there is a clear gap between the top $r$ patterns and the rest.

3. Real-World Uses: Shrinking the System

Once the computer finds these dominant patterns, the authors show how to use them to simplify complex machines (called Linear Time-Invariant Systems).

Model Order Reduction (MOR): Imagine you have a massive, 100-page instruction manual for a robot, but 90 of those pages are just repetitive details. The NPM helps you identify the 10 pages that actually matter. You can then throw away the other 90 pages and create a tiny, 10-page manual that behaves almost exactly like the original. This makes the computer run much faster.
Preserving Stability: The authors proved that when you shrink the system this way, you don't accidentally break it. If the original robot was stable (didn't fall over), the tiny 10-page version will also be stable.

4. Tracking Moving Targets (Time-Varying Systems)

The paper also tested this on systems that change slowly over time, like a drone adjusting its flight path as the wind shifts.

They tried to use the NPM to track the "dominant wind patterns" in real-time.
The Result: It worked very well when the wind changed slowly and predictably.
The Limitation: When the wind changed too quickly (or the mathematical "gap" between the strong and weak patterns was too small), the algorithm got confused and couldn't keep up. The paper notes that for these fast-changing scenarios, the method needs more work to be reliable.

Summary

In short, this paper presents a new, robust way for computers to find the "main characters" in a story of numbers. Unlike older methods that look for specific types of movement, this new method looks for the sheer power of the movement. The authors successfully used this to shrink massive, complex systems into smaller, faster versions without losing their essential behavior, though they admit it struggles when the system changes too rapidly.

Technical Summary: Convergence Analysis of Natural Power Method and Its Applications to Control

Problem Statement
Online principal component algorithms are widely used for extracting dominant subspaces. While the continuous-time Oja flow and its discrete-time Euler approximations are well-studied, they target eigenvalues with the largest real parts. In contrast, the discrete-time Natural Power Method (NPM), defined by the iteration $U[k + 1] = AU[k](U[k]^\top A^\top AU[k])^{-1/2}$ , has primarily been analyzed only for symmetric, positive-definite matrices, where it extracts eigenvalues with the largest magnitudes. A gap exists in the theoretical understanding of the NPM for general square matrices (non-symmetric, indefinite), specifically regarding its convergence properties and its applicability to discrete-time control systems where dominant modes are defined by the largest absolute eigenvalues.

Methodology
The authors conduct a rigorous convergence analysis of the NPM for general square matrices $A \in \mathbb{R}^{n \times n}$ .

Convergence Analysis: The paper generalizes existing results (Hua et al., 1999) to non-symmetric matrices. The authors define the Stiefel manifold $St(r, n)$ and analyze the forward invariant sets of the algorithm. They prove that if the initial condition lies within a specific set $V^{(d)}_{m,r}$ and there is a spectral gap in absolute values ( $|\lambda_m| > |\lambda_{m+1}|$ ), the sequence $U[k]$ converges to an invariant set $U^{(d)}_{m,r}$ corresponding to the $r$ eigenvalues with the largest absolute values.
Stationarity and Modification: The authors demonstrate that the standard NPM iteration generally does not converge to a single stationary point (an element of the invariant set) but rather oscillates within the set. To address this, they propose a modified algorithm (Eq. 9) that includes an additional orthogonal matrix multiplication, ensuring convergence to a specific element of the invariant set.
Dimension Reduction: A reduced-order NPM is proposed to efficiently extract lower-dimensional subspaces from an already computed dominant subspace, avoiding repeated operations on the full $n \times n$ matrix.
Control Applications: The theoretical results are applied to Model Order Reduction (MOR) for discrete-time Linear Time-Invariant (LTI) systems and Low-Rank Controller Synthesis for Linear Time-Varying (LTV) systems. The authors utilize the NPM to construct reduced-order models that preserve observability, reachability, and stability.

Key Contributions and Results

Convergence Characterization: The paper establishes that the NPM extracts the $r$ -dominant subspace corresponding to the $r$ eigenvalues with the largest absolute values. This is a distinct property compared to the Oja flow, which targets the largest real parts.
General Matrix Convergence: The convergence is proven for general square matrices, not just symmetric positive-definite ones. The analysis shows that almost all initial matrices converge to the invariant set, provided a spectral gap exists in the absolute values of the eigenvalues.
Preservation of System Properties: For discrete-time LTI systems, the proposed MOR scheme preserves the minimality (observability and reachability) and stability of the original system. The reduced model shares the same transfer function as the original system projected onto the dominant subspace.
Singular Perturbation and LTV Extension: The method is shown to be applicable to singularly perturbed systems by extracting slow and fast modes. Furthermore, the recursive nature of the NPM allows for its application to slowly varying LTV systems for tracking time-varying dominant subspaces.
Numerical Validation: Simulations using a parameterized matrix $A_\alpha$ demonstrate that the convergence rate depends on the ratio of the absolute values of the eigenvalues ( $|\lambda_{r+1}|/|\lambda_r|$ ). The paper notes that when this ratio is close to 1 (e.g., $\alpha = 0.9$ ), convergence is slow, which impacts the performance of the derived controller in tracking tasks.

Significance and Claims
The authors claim that this work provides a foundational convergence analysis for the NPM in the context of general square matrices, distinguishing it clearly from the Oja flow. The primary significance lies in bridging this theoretical gap to enable practical applications in control theory, specifically for discrete-time systems where the dominant dynamics are governed by the magnitude of eigenvalues rather than their real parts.

The paper modestly frames its contributions as:

Providing a standard convergence benchmark for the NPM on general matrices.
Demonstrating the utility of NPM for Model Order Reduction and low-rank controller synthesis.
Highlighting the potential for tracking time-varying subspaces in LTV systems, while acknowledging that the performance is limited by the convergence speed of the underlying NPM algorithm.

The authors explicitly state that further work is required for rigorous error bound analysis and broader applications to time-varying control and estimation problems, noting that the current controller synthesis for LTV systems is an initial step that performs well only when the NPM convergence is sufficiently fast.

1. The "Biggest" vs. The "Most Real"

2. How It Works (The "Squeeze")

3. Real-World Uses: Shrinking the System

4. Tracking Moving Targets (Time-Varying Systems)

Summary

More like this