🔢 mathematics

A Perturbation Method for Index Detection for Linear Matrix Pencils

This paper establishes rigorous non-asymptotic bounds for the Puiseux expansion of eigenvalues at infinity and estimates the expected eigenvector condition number for randomly perturbed linear matrix pencils, applying these results to the Cayley transform and validating them through numerical simulations.

Original authors: Hanna Blazhko, Michał Wojtylak

Published 2026-03-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanna Blazhko, Michał Wojtylak

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 understand the "health" of a complex machine, like a giant, intricate clockwork system. In the world of mathematics and engineering, these systems are often described by Matrix Pencils. Think of a matrix pencil as a mathematical blueprint that tells you how the machine behaves, especially when it's pushed to its limits (like when a gear spins infinitely fast).

Sometimes, these blueprints have a hidden flaw or a special feature called an "Index."

Index 1 is like a machine that runs smoothly; if you push it, it reacts immediately.
Index 2 is like a machine with a "sticky gear." If you push it, it hesitates for a split second before reacting. This "hesitation" is mathematically represented by a specific block in the blueprint called a Kronecker block.

The problem is that real-world computers are bad at seeing these "sticky gears." When you try to calculate them, the computer's rounding errors usually break the big sticky block into two tiny, separate pieces, making it look like a smooth machine (Index 1) when it's actually a sticky one (Index 2).

This paper proposes a clever way to find out if a machine is truly "sticky" (Index 2) or just looks that way due to computer errors. The authors, Hanna Blazhko and Michał Wojtylak, use two main strategies, which they call Method 1 and Method 2.

The Core Idea: The "Shake" Test

The authors realized that if you gently "shake" the machine (mathematically, add a tiny bit of random noise or perturbation), the way the machine reacts reveals its true nature.

Method 1: The Deterministic "Ruler"

Imagine you have a ruler that tells you exactly how far a spring stretches when you pull it.

The Theory: They proved that if a machine is truly "sticky" (Index 2), and you pull it with a force of size $\tau$ , the reaction (the eigenvalue) will grow at a very specific speed: proportional to the square root of the force ( $\sqrt{\tau}$ ).
The Catch: If the machine is smooth (Index 1), the reaction grows linearly (proportional to $\tau$ ).
The Result: By measuring how fast the reaction grows as you increase the pull, you can tell if the slope on your graph is 1 (smooth) or 0.5 (sticky).
The Limitation: This ruler is very precise, but it requires you to know exactly where the "sticky" parts are hidden inside the machine's blueprint. If the blueprint is messy or scrambled, the ruler might not work perfectly.

Method 2: The Probabilistic "Dice Roll"

Imagine you are trying to find a hidden treasure in a foggy field. Instead of walking a straight line (Method 1), you decide to throw darts randomly.

The Theory: They take the machine's blueprint and add a little bit of random noise (like rolling dice) to it. They do this many times and average the results.
The Magic: Even though the noise is random, the average behavior of the "sticky" machines still follows that special square-root rule ( $\sqrt{\tau}$ ).
The Advantage: This method is much more robust. It doesn't matter if the blueprint is messy or scrambled; the random noise helps "smooth out" the computer's errors, allowing the true "sticky" nature to shine through. It's like using a flashlight in the fog; the random flashes eventually reveal the shape of the object.

The "Cayley Transform": The Magic Mirror

To make these tests work, the authors use a mathematical trick called the Cayley Transform.

Think of this as a magic mirror. If you look at a machine in this mirror, its infinite speed (infinity) gets transformed into a normal, manageable speed.
This allows the computer to handle the "infinite" problems without crashing, while still preserving the "sticky" signature (the Index 2 behavior) so the tests can detect it.

Real-World Examples

The authors tested their ideas on several scenarios:

Toy Models: Simple, made-up machines where they knew the answer. Both methods worked perfectly, clearly showing the difference between smooth and sticky machines.
Vibrating Strings: Imagine a guitar string with a tiny bit of friction. They showed how to detect if the friction creates a "sticky" delay (Index 2) or not.
Electrical Circuits: They analyzed a complex amplifier circuit. Even though the numbers were tiny and messy, their methods correctly identified the circuit's "sticky" nature.

Why Does This Matter?

In engineering, knowing if a system is Index 1 or Index 2 is crucial.

Safety: A "sticky" system might be harder to control. If you design a controller for a smooth machine but the real one is sticky, your controller might fail, causing the machine to crash or oscillate wildly.
Efficiency: Knowing the index helps engineers choose the right mathematical tools to simulate the machine, saving time and computing power.

The Bottom Line

The paper answers three big questions:

Can we see the "sticky" behavior in a computer simulation? Yes, if we look at how the system reacts to tiny changes.
Does this work even when we use digital approximations (discretization)? Yes, thanks to the magic mirror (Cayley transform) and random noise.
Can we use this to measure how close a system is to being "sticky"? Yes! By observing the slope of the reaction, we can tell not just if it's sticky, but how close it is to being perfectly smooth or perfectly sticky.

In short, the authors have built a new, robust "detector" that can spot hidden delays in complex systems, even when the computer is trying to hide them with rounding errors. It's like having X-ray vision for the mathematical blueprints of the machines that run our world.

1. Problem Statement

The paper addresses the challenge of determining the index of a linear matrix pencil $\lambda E - A$ , specifically focusing on the size of the largest Kronecker block associated with the eigenvalue at infinity.

Context: While randomized methods have recently improved the computation of eigenvalues for singular pencils, they often fail to compute the Kronecker canonical form (KCF), which is essential for understanding the system's structural properties (e.g., in energy-based modeling and port-Hamiltonian systems).
The Core Question: Can the fractional Puiseux expansion of eigenvalues at infinity, theoretically predicted by perturbation theory, be observed in numerical simulations and discretizations? If so, can this behavior be used to robustly detect the index (specifically distinguishing between index 1 and index 2)?
Specific Challenges:
1. Numerical algorithms tend to replace a single Kronecker block of size $k$ with $k$ distinct eigenvalues near the singularity, obscuring the theoretical fractional expansion.
2. Determining the index requires distinguishing between the asymptotic behaviors of eigenvalues under perturbation (e.g., $O(\tau^{-1/2})$ for index 2 vs. $O(\tau^{-1})$ for index 1).

2. Methodology

The authors employ a dual approach combining deterministic perturbation theory and probabilistic analysis to analyze the behavior of eigenvalues under small perturbations.

A. Deterministic Approach (Theorem 1.1)

Perturbation Model: The pencil is perturbed as $\lambda(E + \tau I + \Delta E) - (A + \Delta A)$ , where $\tau$ is a scalar parameter and $\Delta E, \Delta A$ are bounded perturbations.
Theoretical Basis: The method relies on the first-order perturbation theory developed in previous works, which states that for an index-2 pencil, the perturbed eigenvalue $\lambda(\tau)$ admits a Puiseux expansion:
$\lambda(\tau) = C \tau^{-1/k} + o(\tau^{-1/k})$
For index 2 ( $k=2$ ), this implies $|\lambda(\tau)| \sim \tau^{-1/2}$ .
Analysis: The authors derive rigorous, non-asymptotic bounds for the magnitude of the perturbed eigenvalue. They utilize the eigenvector condition number ( $\kappa_V$ ) and the Bauer-Fike theorem to bound the deviation of the perturbed eigenvalue from the theoretical expansion.
Key Tool: The analysis involves the Schur complement of the perturbed pencil and specific block decompositions of $E$ and $A$ (specifically $E_{11}, A_{12}, A_{21}$ ).

B. Probabilistic Approach (Theorem 1.2)

Motivation: The deterministic bounds depend on $\kappa_V$ , which can be unbounded or difficult to estimate. To overcome this, the authors introduce random perturbations.
Method: Instead of a deterministic shift $\tau I$ , they use a complex Ginibre matrix (a random matrix with i.i.d. complex Gaussian entries) scaled by $\tau$ .
Cayley Transform: To handle the eigenvalue at infinity, the pencil is transformed via the Cayley transform:
$M_h(E, A) = (E - hA)^{-1}(E + hA)$
This maps the eigenvalue at infinity to a finite value, allowing standard random matrix theory techniques to be applied.
Result: The authors prove that the expected value of the eigenvector condition number for the randomly perturbed matrix is bounded, allowing for a stable estimation of the eigenvalue behavior even in the presence of noise.

C. Special Case: Port-Hamiltonian Systems

The paper specifically tailors these methods to dissipative Hamiltonian pencils ( $\lambda E - (J-R)Q$ ), which arise in energy-based modeling.

They show that for these systems, the eigenvector condition number admits sharper bounds, making the deterministic approach more effective.
They analyze the case where $Q > 0$ , ensuring all eigenvalues lie in the closed left half-plane, with specific structural properties regarding Kronecker blocks at zero or infinity.

3. Key Contributions

Rigorous Non-Asymptotic Bounds: The paper provides explicit upper and lower bounds for the eigenvalues of perturbed index-2 pencils. These bounds confirm that the $O(\tau^{-1/2})$ behavior is observable even with small perturbations $\Delta$ .
Probabilistic Index Detection: A novel method is introduced using random matrix perturbations (Ginibre ensemble) to control the eigenvector condition number. This allows for the detection of the index without requiring precise knowledge of the system's block structure.
Handling Discretization: The authors demonstrate that the Cayley transform preserves the index-detecting features of the pencil, validating the use of discretized models (like the implicit midpoint rule) for index detection.
Distance Estimation: The methods effectively solve a "distance problem," determining how close a given pencil is to the set of index-2 pencils based on the observed convergence rate of eigenvalues.

4. Results

The theoretical findings are validated through extensive numerical simulations on various benchmark examples:

Toy Models: Randomly generated index-2 pencils successfully exhibited the predicted slope of $1/2$ in log-log plots of $1/|\lambda|$ vs. $\tau$ .
Congruence Transformations: The method remained robust even when the pencil was transformed by a random congruence matrix, which obscured the explicit block structure required for the deterministic bounds (though the probabilistic method still worked).
Vibrating Strings (Port-Hamiltonian):
- Case A (Near Index 2): The method correctly identified the $1/2$ slope.
- Case B (Near Index 1): The method correctly identified the $1$ slope, demonstrating clear distinction between indices.
Electrical Circuits: Applied to a CMOS operational amplifier model, the method successfully detected index 2 despite large disparities in matrix entry magnitudes.
Error Analysis: The paper identifies four distinct regions of $\tau$ $τ$ in numerical simulations:
1. Too large $\tau$ : Theoretical asymptotics not yet valid.
2. Optimal $\tau$ : The fractional expansion (slope 1/2) is clearly visible.
3. Suboptimal $\tau$ : Numerical rounding splits the Kronecker block, changing the slope to 1.
4. Too small $\tau$ : Round-off errors dominate.

5. Significance

Bridging Theory and Practice: The paper resolves the question of whether theoretical Puiseux expansions are numerically observable, providing a positive answer with rigorous error bounds.
Robustness: By combining deterministic bounds with probabilistic averaging, the proposed methods are robust against numerical noise and do not require the computationally expensive and unstable calculation of the full Kronecker canonical form.
Applications: The techniques are particularly valuable for energy-based modeling (port-Hamiltonian systems), where knowing the index is crucial for system stability analysis, control design, and ensuring minimality.
Algorithmic Improvement: The work suggests that randomized perturbations can serve as a reliable alternative to the classical "staircase algorithm" for index detection, offering better performance in singular cases.

In conclusion, Blazhko and Wojtylak provide a mathematically rigorous and numerically verified framework for detecting the index of linear matrix pencils, leveraging both perturbation theory and random matrix theory to overcome classical limitations in singular system analysis.