⚛️ high-energy theory

Learning Topological Features of $\widehat Z$ -invariants

This paper presents a systematic machine learning framework that demonstrates neural networks can reliably extract topological information, such as homology classes and graph structures, from $\widehat{Z}$ -invariants of plumbed 3-manifolds while revealing interpretable spectral proxies and uncovering a predictive relationship between these invariants and the Heegaard Floer $d$ -invariant.

Original authors: Brandon Robinson, Shimal Harichurn, Fabian Ruehle, Sergei Gukov, Rak-Kyeong Seong, Miranda C. N. Cheng

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Brandon Robinson, Shimal Harichurn, Fabian Ruehle, Sergei Gukov, Rak-Kyeong Seong, Miranda C. N. Cheng

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

In the quiet corners of mathematics where shapes are studied not by their size or color, but by the way they are connected, there exists a deep mystery about three-dimensional spaces. Imagine a universe that is finite yet has no edges, a closed shape that loops back on itself. Mathematicians have long sought ways to tell these shapes apart, using tools called invariants—numbers or patterns that remain unchanged even if the shape is stretched or twisted. For decades, these tools were like fingerprints: unique to each shape, but difficult to read. Recently, a new kind of tool has emerged from the world of physics and computer science. It involves looking at these shapes through the lens of infinite series, which are essentially endless lists of numbers that follow a specific pattern. These lists, known as q-series, act as a complex code for the geometry of the space. The question researchers have been asking is whether a computer can learn to read this code, not just to guess the shape's identity, but to understand the hidden rules that connect the code to the shape's true nature.

A team of researchers has now taken a systematic approach to this challenge, treating these infinite lists of numbers as a massive dataset to be analyzed by artificial intelligence. They focused on a specific family of three-dimensional shapes built from a network of interconnected loops, which can be described by a simple diagram of dots and lines. For each of these shapes, they generated a corresponding list of numbers, truncating the infinite series to the first ten thousand terms to make it manageable for a computer. This resulted in a database of over sixty thousand distinct examples. The researchers then trained neural networks, a type of artificial intelligence designed to recognize patterns, to perform two main tasks. First, they asked the computer to distinguish between shapes that are topologically equivalent to a perfect sphere and those that are not. Second, they asked the computer to identify the underlying structure of the diagram used to build the shape, such as whether it looked like a star or an H.

The results were striking. The neural networks learned to identify these topological features with extremely high accuracy, often reaching nearly perfect scores. However, the true breakthrough was not just that the computer got the right answers, but that the researchers could figure out how it was doing it. By peering inside the decision-making process of the network, they discovered that the computer was not blindly memorizing the endless lists of numbers. Instead, it had learned to extract specific, meaningful features from the data. For the task of identifying sphere-like shapes, the network relied heavily on a specific statistical property of the shape's underlying diagram: the size of its smallest eigenvalue, a number that describes how the shape's structure vibrates or stretches. This was a clever shortcut. The researchers found that for shapes that are true spheres, this number is always very small, providing a reliable signal that the computer could use to make its decision.

In another set of experiments, the team explored a different kind of relationship, one that connects these geometric shapes to a concept called homology cobordism. This is a way of determining if two shapes can be smoothly transformed into one another through a higher-dimensional space. They trained the computer to predict a specific correction term, a number that helps mathematicians classify these shapes, using only the leading numbers from the infinite series. Surprisingly, the computer learned this relationship with over ninety-nine percent accuracy. When they added a few extra mathematical details derived from the shape's construction to the training data, the computer's predictions became nearly perfect, correctly identifying the integer value in almost ninety percent of cases. This suggests that the infinite series contains a hidden, subtle geometric message about how these shapes relate to one another, a message that was previously difficult to decode.

The study also revealed a fascinating limitation in how the computer learned. While it was highly accurate, it sometimes relied on "surrogate" features—patterns that were correlated with the correct answer but were not the exact mathematical rule. For instance, in some cases, the network used the spacing between the non-zero numbers in the list as a proxy for the shape's structure. This happened because the network found a statistical shortcut that worked for the vast majority of the data it saw, even if it wasn't the fundamental law. This behavior highlights a crucial insight: the computer is not simply a black box that guesses; it is a tool that can reveal which mathematical properties are most visible in the data. By showing us that the network prioritizes certain spectral properties over others, the researchers have provided a new way to understand these complex shapes.

Ultimately, this work demonstrates that machine learning can serve as a powerful microscope for pure mathematics. It allows researchers to sift through vast amounts of abstract data to find the specific features that define a shape's identity. The findings suggest that the infinite series used to describe these three-dimensional spaces carry a rich amount of geometric information, including details about how the shapes can be transformed into one another. While the computer's methods are sometimes different from the traditional proofs used by mathematicians, the results point toward new connections between different areas of mathematics. The study opens a door for future research, suggesting that by teaching computers to read these mathematical codes, we may uncover new relationships and conjectures that have remained hidden for decades.

Technical Summary: Learning Topological Features of $\widehat{Z}$ -invariants

Problem Statement
The paper addresses the challenge of extracting interpretable topological information from $\widehat{Z}$ -invariants (homological blocks) of closed 3-manifolds. These invariants are defined as infinite $q$ -series with integer coefficients, arising from M-theory and related to quantum modular forms. While $\widehat{Z}$ -invariants are known to encode deep topological data, the specific mechanisms by which they capture properties such as homology classes, graph structures, and cobordism invariants remain partially mysterious. The authors aim to systematically analyze these truncated infinite series using machine learning (ML) to identify patterns, formulate conjectures, and understand the "black box" decision-making processes of neural networks in a mathematical context.

Methodology
The authors construct a comprehensive data analysis pipeline involving three main stages: data generation, representation, and interpretability.

Data Generation: They generate a dataset of approximately $6.3 \times 10^4$ truncated $\widehat{Z}$ -invariants (up to $10^4$ terms) for plumbed 3-manifolds. The manifolds are derived from two primary graph topologies: star-shaped graphs (corresponding to Seifert fibered spaces) and H-shaped graphs. The dataset includes both integral homology spheres (where the plumbing matrix determinant is $\pm 1$ ) and non-spheres.
Data Representations: To ensure robustness, the authors test three input representations:
- Full Sequence: A fixed-length array of coefficients $[c_0, \dots, c_{N-1}]$ .
- Value-Index Pair: A compressed representation storing only non-zero coefficients as tuples $(c_i, n_i)$ , explicitly providing exponent indices.
- Latent Space: Low-dimensional embeddings (5D) generated via contrastive learning (semi-hard triplet loss) to filter high-variance noise.
Interpretability Pipeline: A core methodological innovation is the use of a "teacher-student" framework for interpretability.
- A primary classifier (teacher) is trained on the raw $q$ -series data to perform topological classification (e.g., sphere vs. non-sphere, star vs. H-graph).
- A secondary regression network (student) is trained to map a set of ten pre-defined, mathematically motivated abstract features (e.g., plumbing matrix eigenvalues, determinant, gap statistics of the $q$ -series) to the teacher's output logits.
- Feature importance is assessed using Feature Saliency (FS), Permutation Importance (PI), and Shapley Additive Explanations (SHAP) to determine which abstract features drive the teacher's decisions.

Key Contributions and Results

Topological Classification: Neural networks can reliably classify homology spheres and distinguish graph topologies directly from $\widehat{Z}$ -series coefficients with high balanced accuracy (often $>95\%$ ).
- Homology Classification: The networks learn to distinguish spheres from non-spheres. Interpretability analysis reveals that the networks often rely on spectral proxies rather than the determinant directly. Specifically, the minimum magnitude eigenvalue ( $\epsilon_{\min}$ ) of the plumbing matrix serves as a strong proxy for the determinant condition $|\det M|=1$ , supported by a spectral bound derived in the paper (Lemma 1).
- Graph Topology Classification: Distinguishing star-shaped from H-shaped graphs depends heavily on the homology class. For homology spheres (where the determinant is fixed), the decision boundary is driven by gap statistics (mean and standard deviation of the exponents of non-zero coefficients, $\mu_\gamma, \sigma_\gamma$ ). For non-spheres, the decision combines determinant information with spectral data.
- Representation Impact: The Value-Index Pair representation yields the highest performance and the most stable feature-to-logit reconstruction, suggesting that explicitly providing exponent indices allows the network to efficiently utilize the sparsity structure of the series.
Cobordism and the $d$ -invariant: The paper investigates the relationship between $\widehat{Z}$ -invariants and the Heegaard Floer correction term ( $d$ -invariant), a homology cobordism invariant.
- On datasets of Seifert integer homology spheres (specifically Brieskorn spheres with 3 singular fibers), regression models can predict the $d$ -invariant (or equivalently the $\tau$ -function contribution) from the leading exponents of the $\widehat{Z}$ -series with $>99.9\%$ explained variance ( $R^2$ ).
- The relationship is effectively captured by a quadratic polynomial of the leading exponent offsets.
- By augmenting the input with symmetrized Dedekind sums derived from the Seifert invariants, the model achieves nearly 90% accuracy in predicting the exact integer value of $d$ .
- The authors note that while the $q$ -series exponents contain sufficient arithmetic information to reconstruct the Seifert parameters (and thus $d$ ) in the Brieskorn case, this does not generalize to out-of-distribution families (e.g., non-square-free Brieskorn spheres or general manifolds), where the models fail.
Interpretability Insights: The study demonstrates that ML models do not merely find opaque correlations but often learn specific, mathematically justifiable proxies. For instance, the reliance on $\epsilon_{\min}$ is explained by a rigorous spectral bound. The "surrogate" features learned by the network (e.g., gap statistics) often provide a more stable signal than the exact topological definitions in finite datasets, explaining both the high accuracy and the specific generalization errors observed.

Significance and Claims
The paper claims to establish a systematic framework for handling mathematical data structured as truncated infinite $q$ -series. Its primary significance lies in:

Methodological Advancement: Demonstrating that ML can extract human-understandable topological rules from complex series data, moving beyond simple prediction to mechanistic understanding via interpretability regression.
Novel Connections: Revealing a strong, high-accuracy predictive relationship between the exponents of $\widehat{Z}$ -invariants and the Heegaard Floer $d$ -invariant for Seifert homology spheres. This suggests that $\widehat{Z}$ -invariants capture subtle geometric information regarding homology cobordism equivalences.
Cautionary Note: The authors modestly frame their findings, noting that while the models learn robust interpolation rules within specific families (like Brieskorn spheres), they do not necessarily learn general closed-form formulas for topological invariants (as evidenced by the failure on out-of-distribution $SL_2(\mathbb{Z})$ orbit sizes).

The work suggests a new direction for studying quantum invariants, where machine learning serves as a tool to identify which spectral and geometric proxies are most relevant to topological classification and cobordism properties.

More like this