🔬 mesoscale physics

Machine learning prediction of the convergence criterion for a topological invariant of finite non-Hermitian chains

This paper demonstrates that machine learning, specifically random-forest regression, can accurately predict the optimal crop-length parameter for calculating topological invariants in finite non-Hermitian chains by linking it to physical decay lengths and characteristic polynomial structures, thereby providing a robust, generalizable, and disorder-resilient method to capture topology near phase transitions.

Original authors: Raghav Chaturvedi, Viktor Könye, Ewelina M. Hankiewicz

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Raghav Chaturvedi, Viktor Könye, Ewelina M. Hankiewicz

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: Measuring a Wobbly Rope

Imagine you have a long, wobbly rope (a quantum system) that is tied down at both ends. This rope has a special property: it can twist and loop in invisible ways that we call "topology." In a perfect, infinite world, we could easily count how many times the rope twists.

However, in the real world, our rope is finite (it has a specific length) and it's wobbly (it's "non-Hermitian," meaning energy can leak in or out, like a balloon slowly deflating). Because of this wobble, the ends of the rope get messy and chaotic. If you try to count the twists right at the ends, you get the wrong answer.

To get the right count, you have to ignore the messy ends and only look at the clean middle section of the rope. The paper asks a simple question: How much of the messy end do we need to cut off (or "crop") to see the true pattern?

The authors call this the "crop-length." If you cut off too little, the mess ruins your count. If you cut off too much, you might cut into the clean part and lose data. The goal is to find the perfect amount to cut.

The Problem: Guessing the Cut

Usually, scientists just guess how much to cut off. They try a little, then a little more, until the number stops changing. This is slow and tricky, especially when the rope is about to change its shape (a "phase transition").

The authors wanted to build a smart assistant (a machine learning model) that could look at the rope's properties and instantly tell you exactly how much to cut off to get the perfect answer.

The Secret Ingredient: The "Decay Length"

The paper discovered that the amount you need to cut off isn't random. It is controlled by a physical property called the "localization length" (or decay length).

The Analogy: Imagine the messy ends of the rope are like a stain spreading from the edge. The "decay length" is how far that stain spreads before it fades away.
The Rule: To get a clean reading, you must cut off a distance roughly equal to the size of that stain. If the stain spreads 10 inches, you need to cut off at least 10 inches.

The authors found that for simple ropes (where the wobble only happens between immediate neighbors), there is only one stain size. The machine learning model learned that the "crop-length" is just a simple multiple of this stain size.

The Machine Learning Magic

The researchers trained a computer program (a Random Forest model) to act as a detective.

The Input: They fed the computer the details of the rope (its length, how wobbly it is, and how fast the stain spreads).
The Output: The computer predicted the exact number of inches to cut off.

The Results:

Simple Ropes: For basic ropes, the computer was almost perfect. It realized that the "stain size" was the most important clue.
Complex Ropes: When they made the ropes more complicated (where the wobble jumps over neighbors or happens in multiple directions), the "stain" wasn't just one size anymore. It became a mix of different spreading rates. The computer learned to look at the slowest spreading stain (the one that goes the furthest) to decide how much to cut.
Generalization: The best part? They trained the computer on one type of rope, and it could accurately predict the cut for completely different ropes it had never seen before, including ones with different colors (complex energies) and different lengths.

What About Messy Ropes? (Disorder)

Finally, they tested if this smart assistant would still work if the rope itself was a bit broken or uneven (disorder).

The Finding: As long as the rope wasn't too broken, the assistant's prediction still worked well in the middle of the rope.
The Limit: Near the very edges of the "clean" zones (where the rope is about to change shape), the prediction got a bit shaky. This makes sense because that's where the "stain" is most sensitive to changes.

Summary

This paper teaches us that to measure the hidden twists of a finite, wobbly quantum rope, you don't need to guess. You just need to know how far the "mess" spreads from the edges.

The authors built a machine learning tool that looks at the physics of the mess and tells you exactly how much of the edge to cut off to get a perfect measurement. This works for simple ropes, complex ropes, and even slightly broken ropes, making it a practical guide for scientists studying these strange quantum systems.

Technical Summary: Machine Learning Prediction of the Convergence Criterion for a Topological Invariant of Finite Non-Hermitian Chains

Problem Statement
In non-Hermitian quantum systems, topology is often characterized by point gaps in the complex energy spectrum, leading to phenomena such as the non-Hermitian skin effect (NHSE), where extensive numbers of eigenstates localize at boundaries. While momentum-space winding numbers ( $w$ ) define topology in translationally invariant systems, finite open chains require real-space topological invariants. The polar-decomposition real-space invariant ( $w_{PD}$ ) is a primary candidate for this purpose. However, $w_{PD}$ depends on a "crop-length" parameter ( $\ell$ ), which defines a central bulk window by excluding boundary regions. The value of $\ell$ required for $w_{PD}$ to converge to the exact momentum-space winding number is typically chosen empirically. This choice becomes critical near topological phase transitions where finite-size effects are strongest, yet a systematic, physically grounded method to determine the optimal crop-length ( $\ell^\star$ ) has been lacking.

Methodology
The authors frame the prediction of the minimum required crop-length $\ell^\star$ (defined as the smallest $\ell$ such that $|w_{PD}(\ell) - w| < \epsilon$ ) as a supervised machine learning problem.

Data Generation: They generate datasets for various finite non-Hermitian chains, including nearest-neighbor (Hatano-Nelson), pure $m$ -hop, and mixed-range hopping models. Samples are defined by system size ( $N$ ), hopping parameters, and base energy ( $E_B$ ).
Feature Engineering: The input features are derived from the physical properties of the systems. Crucially, they utilize the roots ( $\beta$ ) of the characteristic polynomial $H(\beta) - E_B = 0$ . The modulus of these roots determines the spatial decay lengths ( $\xi$ ) of the eigenstates. Features include raw hopping parameters, non-reciprocity measures, and derived decay lengths or grouped root channels (accounting for complex conjugate pairs).
Models: Random-forest (RF) models are employed. A classifier determines if a valid crop exists within the allowed range, while a regressor predicts the numerical value of $\ell^\star$ for valid samples. The models are trained and tested using an 80/20 split, with specific attention to separating finite-size residue classes (e.g., $N \mod 2m$ ) to ensure generalization.
Validation: The approach is tested across varying hopping ranges ( $m=1$ to $5$), complex base energies ( $E_B$ ), and under moderate hopping disorder.

Key Results

Nearest-Neighbor and Pure $m$ -Hop Chains: For Hatano-Nelson chains with nearest-neighbor ( $m=1$ ) or pure $m$ -hop hopping, the required crop-length is controlled by a single localization length scale ( $\xi$ or $\xi_m$ ).
- The random forest regressor achieves high accuracy ( $R^2 \approx 0.997$ ) in predicting $\ell^\star$ .
- Feature importance analysis confirms that the scaling variable $x = N/\xi$ is the dominant predictor.
- Away from phase transitions, $\ell^\star$ scales linearly with the localization length ( $\ell^\star \propto \xi$ ). Near transitions, finite-size corrections are captured by a rational function of $x$ .
- The tolerance dependence follows $\ell^\star \approx C_\epsilon \xi$ , where $C_\epsilon$ grows linearly with $\log(1/\epsilon)$ , consistent with exponentially decaying boundary contributions.
Mixed Hopping Models: For chains with mixed first- and second-neighbor hoppings, the spectral loops are non-elliptic, and the characteristic equation yields multiple distinct decay channels.
- A single localization length is insufficient. Instead, the crop-length is governed by the "grouped channel" closest to the unit circle in the complex $\beta$ -plane (the slowest decaying mode).
- Random forest models trained on grouped root-channel features successfully predict $\ell^\star$ across different winding sectors ( $w = \pm 1, \pm 2$ ) with $R^2 > 0.98$ .
Complex Base Energies: The method extends to general complex base energies $E_B$ .
- For pure $m$ -hop models, the root structure remains constrained (two families of roots), and the predictor generalizes to unseen hopping parameters via leave-one-out validation ( $R^2 = 0.993$ ).
- For mixed hopping models with complex $E_B$ , the predictor requires the full signed radial root vector (distance and inside/outside status relative to the unit circle) to distinguish nested topological sectors. This representation allows accurate reconstruction of $w_{PD}$ maps across the complex energy plane for unseen Hamiltonians.
Robustness to Disorder: The predictor trained on clean nearest-neighbor chains remains effective under moderate hopping disorder ( $r_W = 0.3$ ).
- The predicted $\ell^\star$ from the clean system yields accurate $w_{PD}$ values in the interior of topological phases (success fraction $> 0.96$ ).
- Performance degrades near phase boundaries where the point gap closes, consistent with increased sensitivity to finite-size effects and disorder.

Significance and Claims
The paper claims to provide a "practical and physically interpretable way to choose the crop-length" for real-space topological invariants in finite non-Hermitian systems. By demonstrating that machine learning can accurately generalize the relationship between physical decay lengths and the required crop-length across diverse Hamiltonians, the authors establish that the convergence of $w_{PD}$ is fundamentally controlled by the localization lengths associated with the non-Hermitian skin effect.

The authors emphasize that their results allow for the reliable determination of topology in finite systems without arbitrary parameter tuning. Furthermore, they suggest that the learned relationships can be inverted to estimate effective skin-effect decay lengths from experimental or numerical data, even in the presence of disorder, provided the system is not too close to a phase boundary. The study does not propose new experimental platforms but rather offers a computational tool to interpret existing or future data on non-Hermitian topological phases.