📊 statistics

Quantum Geometry of Data

This paper establishes Quantum Cognition Machine Learning (QCML) as a foundational framework that encodes data into quantum geometry via learned Hermitian matrices, enabling the extraction of global topological invariants and efficient, interpretable representations of high-dimensional datasets without succumbing to the curse of dimensionality.

Original authors: Alexander G. Abanov, Luca Candelori, Harold C. Steinacker, Martin T. Wells, Jerome R. Busemeyer, Cameron J. Hogan, Vahagn Kirakosyan, Nicola Marzari, Sunil Pinnamaneni, Dario Villani, Mengjia Xu, Khar

Published 2026-09-30

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexander G. Abanov, Luca Candelori, Harold C. Steinacker, Martin T. Wells, Jerome R. Busemeyer, Cameron J. Hogan, Vahagn Kirakosyan, Nicola Marzari, Sunil Pinnamaneni, Dario Villani, Mengjia Xu, Kharen Musaelian

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 modern world, data is often overwhelming. A single patient record might contain dozens of measurements, from blood pressure and heart rate to complex genetic sequences and medical images. When scientists try to find patterns in such vast collections, they face a fundamental problem: as the number of features grows, the amount of data needed to understand them grows exponentially, quickly becoming impossible to manage. This is known as the curse of dimensionality. To solve this, researchers have long tried to find the hidden, simpler shapes that data points actually form, much like realizing that a cloud of scattered dust in a room is actually arranged along a thin, invisible wire. Traditional methods often look at how close individual points are to their neighbors, building a map based on local connections. However, this approach can miss the bigger picture, failing to see the global structure or the deep, topological twists that define the entire dataset.

A new approach called Quantum Cognition Machine Learning offers a different way to see these patterns. Instead of just measuring distances between points, this method treats the data as if it were a quantum system. In this framework, every data point is translated into a specific state, and the features that describe that point are represented by mathematical objects that act like observables in physics. By training a computer to find the best arrangement of these objects, the method creates a compact, geometric model of the data. This model is not just a list of numbers; it is a rich, multi-dimensional shape that captures the essential nature of the information, revealing properties like curvature and connectivity that are invisible to standard techniques.

In a recent study, a team of researchers demonstrated how this quantum-inspired framework can be used to extract the true geometric and topological structure of real-world data. They showed that even highly complex datasets, containing dozens of different features, could be faithfully represented in a space that is surprisingly small. For instance, they found that a dataset with thirty different measurements could be accurately modeled using a mathematical space with a dimension as small as eight. This is a remarkable compression, suggesting that the essential information in these complex systems is far more organized than it appears on the surface. The researchers did not just build a model; they developed a way to read the geometry of that model, uncovering hidden shapes, counting the number of separate pieces in the data, and identifying the specific features that hold the most importance.

The team tested their method on several types of data, starting with synthetic examples where the answer was already known. They created datasets that looked like two separate spheres floating in space, and others that resembled a single sphere with points clustered more densely in one area. In every case, the quantum geometry learned by the machine perfectly matched the underlying shape. It could distinguish between the two separate spheres and identify the single sphere as one continuous object, even when the data was noisy. The method also successfully detected the "holes" or twists in the data, measuring topological charges that act like fingerprints for the shape's global structure. This ability to see the whole picture, rather than just the local neighborhoods, allowed the researchers to identify the intrinsic dimension of the data—the true number of directions in which the information varies—without needing to guess or rely on arbitrary rules.

One of the most significant tests involved a real-world dataset containing medical records of breast cancer patients. This dataset included 569 samples, each described by thirty different features derived from images of cell nuclei. The goal was to see if the quantum geometry could distinguish between benign and malignant tumors without being told which was which during the training process. The researchers found that the method naturally separated the two groups into distinct regions of the geometric space. The malignant samples clustered together, and the benign samples formed their own group, creating a clear boundary between them. This separation emerged purely from the geometric structure of the data, revealing that the differences between healthy and cancerous cells are encoded in a way that this method can visualize.

The study also explored how the method handles the trade-off between detail and simplicity. By adjusting a single parameter, the researchers could change the resolution of the geometric model. At high resolution, the model showed fine details and a higher number of dimensions, capturing subtle variations in the data. At lower resolution, the model smoothed out the noise, revealing a simpler, more robust structure. In the breast cancer dataset, this allowed the team to see that the data could be understood as a three-dimensional object, a finding that aligned with other statistical methods but was derived through a completely different, geometric lens. Furthermore, the researchers could identify which of the thirty original features were most important for defining this shape. They found that features related to the size and irregularity of the cell nuclei were the primary drivers of the geometric separation, mirroring the biological reality that cancerous cells tend to be larger and more irregular than healthy ones.

What makes this work particularly powerful is that it does not rely on the data being perfectly smooth or following a simple linear path. The method is designed to handle the messy, non-linear reality of scientific data. It treats the dataset as a collection of quantum cells, where the uncertainty in the measurements is not a flaw to be ignored but a feature that helps define the shape. This uncertainty acts as a natural filter, preventing the model from overfitting to random noise and ensuring that the resulting geometry reflects the true underlying structure. The researchers showed that this approach could capture global properties, such as whether the data forms one connected piece or several disconnected islands, and could even detect the presence of "monopoles," which are points where the geometry becomes singular or degenerate.

The implications of this work extend beyond just finding better ways to classify tumors. The researchers established a new foundation for thinking about data as a geometric object that can be studied with the tools of physics. They demonstrated that it is possible to extract topological invariants, which are numbers that describe the global shape of the data and remain unchanged even if the data is stretched or twisted. These invariants provide a level of understanding that is difficult to achieve with traditional methods. For example, the team showed that the data from the breast cancer study could be mapped onto a space where the separation between benign and malignant cases was geometrically clear, offering a new way to interpret the relevance of different features.

The study also highlighted the efficiency of this approach. While traditional methods often require a large number of components to explain the variance in a dataset, this quantum geometric method achieved similar or better results with far fewer dimensions. In the case of the breast cancer data, a model with a dimension of eight was sufficient to capture the essential structure, whereas other methods might require many more components to reach the same level of understanding. This efficiency suggests that the method can scale to even larger and more complex datasets, such as those found in genomics or climate science, where the number of features can be in the thousands.

Ultimately, this research provides a new language for describing data. It moves beyond simple lists of numbers and distances, offering a way to visualize the shape of information itself. By learning the geometry of the data, the method reveals the hidden organization that governs complex systems. The researchers showed that this approach is not just a theoretical exercise but a practical tool that can be applied to real-world problems. Whether it is distinguishing between different types of cells, understanding the structure of conformal maps, or analyzing the connectivity of a dataset, the quantum geometric perspective offers a fresh and powerful way to see the world. The work suggests that by treating data as a quantum object, we can uncover patterns and structures that have remained hidden, providing a deeper and more intuitive understanding of the information that shapes our world.

Technical Summary: Quantum Geometry of Data

Problem Statement

Data analysis frequently involves high-dimensional datasets (e.g., medical records with vitals, RNA sequencing, and images) where the number of features ( $D$ ) is large. Classical approaches face the "curse of dimensionality," requiring an exponential number of data points to sample the distribution effectively. While manifold learning methods (e.g., Laplacian eigenmaps, diffusion maps) attempt to uncover low-dimensional structures, they typically rely on local neighborhood structures and struggle to capture global geometric and topological properties. Furthermore, linear techniques like Principal Component Analysis (PCA) often require many components to capture essential nonlinear structures, failing to provide a compact representation of the data's intrinsic geometry.

Methodology: Quantum Cognition Machine Learning (QCML)

The paper introduces Quantum Cognition Machine Learning (QCML) as a framework to encode data as quantum geometry. Rather than treating data points as vectors in Euclidean space, QCML maps them to states in a finite-dimensional Hilbert space ( $N$ -dimensional) and represents features as learned Hermitian matrices (observables).

Core Framework

Representation:
- A dataset $X = \{x_1, \dots, x_T\}$ with $D$ features is mapped to a set of $N \times N$ Hermitian matrices $\mathbf{X} = \{X_1, \dots, X_D\}$ .
- Each data point $x_t$ is associated with a quasi-coherent state $|x_t\rangle$ , defined as the ground state (lowest eigenvalue) of a displacement Hamiltonian:
  $H(x) = \frac{1}{2} \sum_{a=1}^D (X_a - x_a I_N)^2$
- The state $|x_t\rangle$ allows for the evaluation of all $D$ features via expectation values: $X(x) = (\langle x|X_1|x\rangle, \dots, \langle x|X_D|x\rangle)$ .
Training Objective:
- The matrices $\mathbf{X}$ are optimized to minimize a loss function $L[\mathbf{X}]$ that balances the fidelity of the data reconstruction against quantum fluctuations:
  $L[\mathbf{X}] = \sum_{x \in X} \left( d^2(x) + w \cdot \sigma^2(x) \right)$
  where $d^2(x)$ is the squared displacement between the original point and its quantum image, and $\sigma^2(x)$ is the variance (quantum uncertainty) of the observables.
- The hyperparameter $w$ controls the "resolution" or coarse-graining. Small $w$ allows for non-commutative (quantum) geometries, while large $w$ forces commutativity, effectively reducing the model to $N$ -means clustering.
Quantum Geometric Tools:
Once the matrices are learned, the framework applies tools from quantum geometry and noncommutative geometry to extract structural properties:
- Quantum Metric and Berry Curvature: Derived from the inner product of quasi-coherent states, these define the intrinsic distance and local uncertainty on the emergent manifold.
- Topological Invariants: The framework computes Chern numbers (integrals of the Berry curvature over 2-cycles) to identify topological charges and monopoles (degeneracy points of the Hamiltonian).
- Matrix Laplacian: Defined as $\Delta = \sum_a [X_a, [X_a, \cdot]]$ $Δ = \sum_{a} [X_{a}, [X_{a}, \cdot]]$ , this operator acts on the space of matrices. Its spectrum and eigenmaps (eigenvectors of the Laplacian) are used to:
  - Estimate intrinsic dimension via Weyl's law.
  - Identify disconnected components (zero modes).
  - Construct Laplacian eigenmaps for dimensionality reduction that preserve global geometric structure.

Key Contributions

The paper establishes QCML as a general framework for matrix geometry (or fuzzy geometry), bridging data science with mathematical physics. Its primary contributions are:

Formulation of QCML as Matrix Geometry: It provides the first comprehensive exposition of QCML as a framework where data points correspond to quasi-coherent states and features to Hermitian operators, defining a noncommutative space.
Extraction of Geometric and Topological Invariants: The work demonstrates how to compute fundamental quantities—quantum geometric tensors, Berry curvature, and Chern numbers—directly from learned data representations.
Introduction of the Matrix Laplacian: It introduces the matrix Laplacian and its eigenmaps as a geometry-preserving alternative to graph-based embeddings, capable of revealing global structure and intrinsic dimension.
Efficiency and Scalability: The paper demonstrates that high-dimensional datasets can be faithfully modeled using Hilbert spaces of remarkably small dimension ( $N \sim 10$ ), overcoming the curse of dimensionality by capturing global nonlinear structures with few degrees of freedom.
Unified Paradigm: It unifies concepts from quantum geometry and data science, offering a new operator-based paradigm for analyzing data structure, including connectivity, curvature, and topological defects.

Results and Examples

The authors validate the framework on both synthetic and real-world datasets:

Synthetic Geometric Objects:
- Fuzzy Sphere: Training on points uniformly distributed on a sphere successfully recovered the geometry of a "fuzzy sphere," with learned operators approximating angular momentum generators and a spectrum matching the theoretical fuzzy sphere Laplacian.
- Disconnected Spheres: The model correctly identified two disconnected spherical components and a "bridge" formed by noise, detecting seven degeneracy points (monopoles) with topological charges $\pm 1$ . The matrix Laplacian spectrum revealed the two-component structure via near-zero eigenvalues.
- Non-uniform Distribution: On a sphere with non-uniform sampling, the model filled in sparse regions with coherent points while maintaining high quantum uncertainty in low-density areas, preserving the underlying topology.
- Conformal Maps: For a dataset of conformal maps (intrinsic dimension $d=2$ embedded in $D=200$ ), QCML correctly identified the intrinsic dimension via the quantum metric spectrum and Weyl's law. The Laplacian eigenmaps successfully recovered the parameters of the maps, demonstrating the ability to learn symmetries and topology without inductive biases.
Real-World Data (Wisconsin Breast Cancer):
- Applied to the WBC dataset ( $D=30, T=569$ ), QCML estimated the intrinsic dimension to be between 2 and 7 depending on the resolution parameter $w$ .
- Feature Relevance: The overlap between learned observables and Laplacian eigenmaps provided an interpretable decomposition of feature contributions, distinguishing global shape features from local texture.
- Separability: The unsupervised eigenmap embedding ( $Y_1, Y_2, Y_3$ ) showed visible separation between benign and malignant samples.
- Classification Performance: Using 3 eigenmap coordinates, logistic regression achieved an ROC AUC of 0.960. While this was lower than PCA (0.990) on the same dataset, the authors note this is expected because the WBC dataset's diagnostic signal is largely linear and low-dimensional. The value of QCML here lies in the geometric interpretation and feature decomposition rather than raw classification accuracy.

Significance and Claims

The paper explicitly positions its contribution as foundational rather than algorithmic benchmarking.

New Representation, Not Just a New Algorithm: The authors state that the primary goal is not to outperform existing classifiers on standard tasks but to develop a new representation of data. This representation maps high-dimensional datasets to finite-dimensional quantum (matrix) geometries.
Access to Global Structure: The significance lies in the ability to extract geometric and topological structures (intrinsic dimension, curvature, Chern numbers) that are not naturally accessible via conventional graph-based or linear frameworks (like PCA).
Efficiency as Compactness: "Efficiency" is defined as representational compactness. The framework demonstrates that complex, high-dimensional data can be captured with small Hilbert space dimensions ( $N \sim 10$ ), revealing global structures that linear methods miss.
Modesty on Performance: The authors are modest regarding immediate performance advantages. They acknowledge that for datasets with linearly separable classes (like WBC), classical methods may perform better. The advantage of QCML is posited for datasets with nontrivial topology, symmetry, or complex global structures where linear methods fail.
Future Directions: The work establishes a bridge for future research in applying quantum geometric tools (spectral analysis, topological invariants) to data science, particularly in domains where topology and uncertainty are central. It also notes that QCML is naturally suited for implementation on future quantum computers.

In summary, the paper argues that viewing data through the lens of quantum geometry provides a robust, scalable, and interpretable framework for understanding the global geometric and topological organization of high-dimensional datasets.