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
Organizations: Qognitive, Inc., Miami Beach, FL 33139, USA · Stony Brook University, Department of Physics and Astronomy, Stony Brook, NY 11790, USA · Wayne State University, Department of Mathematics, Detroit, MI 48202, USA · University of Vienna, Faculty of Physics, Boltzmanngasse 5, A-1090 Vienna, Austria · Cornell University, Department of Statistics and Data Science, Ithaca, NY 14853, USA · Department of Psychological and Brain Sciences, Indiana University. · Theory and Simulations of Materials (THEOS), and National Centre for Computational Design and Discovery of Novel Materials (MARVEL), ´Ecole Polytechnique F´ed´erale de Lausanne, 1015 Lausanne, Switzerland · King’s College London, Department of Mathematics, London, UK · New Jersey Institute of Technology, Department of Data Science, Newark, NJ 07102, USA · Massachusetts Institute of Technology, Center for Brains, Minds and Machines, Cambridge, MA 02139, USA
We demonstrate how Quantum Cognition Machine Learning (QCML) encodes data as quantum geometry. In QCML, features of the data are represented by learned Hermitian matrices, and data points are mapped to states in Hilbert space. The quantum geometry description endows the dataset with rich geometric and topological structure---including intrinsic dimension, quantum metric, and Berry curvature---derived directly from the data. QCML captures global properties of data through a compact operator representation. At fixed Hilbert-space dimension, the number of operator parameters grows only linearly with the number of features. The present work provides the first full exposition of QCML as a framework for matrix geometry, establishing its mathematical foundation and demonstrating how operator-level tools from quantum geometry can be directly applied to data. We introduce the matrix Laplacian and its eigenmaps as a geometry-preserving alternative to graph-based embeddings, and we show how topological invariants such as Chern numbers can be extracted from data for the first time. An important practical outcome is the efficiency of the representation: the relevant geometric structure of high-dimensional datasets can be captured with Hilbert spaces of size as small as N∼10. We illustrate these results on a number of synthetic and real-world examples, including the Wisconsin Breast Cancer dataset, where QCML uncovers geometric separability between benign and malignant samples and yields an interpretable decomposition of feature relevance. The emphasis of this work is on a new geometric representation of data rather than on algorithmic benchmarking, establishing a foundation for extracting global geometric and topological structure from high-dimensional datasets.
Figures & tables
Figure 1: Conceptual map placing this work within the context of geometric data analysis and manifold approximations. The top part of the diagram illustrates two established approaches to approximating a smooth manifold (a): via classical graph-based discretizations (b), such as triangulations, or via quantum or fuzzy geometries (c), as developed in mathematical physics. The bottom part of the diagram represents the inverse problem—recovering geometry from data. A dataset (d), often sampled from a low-dimensional manifold, can be analyzed using graph-based manifold learning methods (b) (bottom-left blue arrow), or alternatively, via the recently introduced QCML approach (c) (bottom-right red arrow). QCML constructs a quantum geometry from data using learned observables and quasi-coherent states. This quantum geometric approach is the central focus of this work.
Figure 2: Learned geometric representations from trained quantum operators X . The unit sphere is shown for reference. Left: The original data, consisting of 1000 points generated on the surface of a unit sphere with a uniform measure, is shown in black. Right: Quantum-learned QCML point cloud (green) with degeneracy points (blue). There are three nearby degeneracy points, each with topological charge −1 , separated by a distance of about 0.03. The training was performed with hyperparameters N=4 , w=0.1 .
Figure 3: Schematic diagram illustrating QCML procedure. Observables Xa (quantum geometry) and quasi-coherent states ∣xt⟩ are obtained through the training using data dependent displacement Hamiltonians and the loss function optimizing a combination of displacement and variance of the data. The learned operators and states are then used to obtain quantum-geometric characteristics of the data. Then these characteristics can be used for extracting intrinsic dimension, clustering, topological properties, inference (expectation values), etc.
Figure 4: Comparison between the spectrum of the matrix Laplacian learned from the data shown in Figure 2 and the exact spectrum of the fuzzy sphere S42 . The blue points correspond to the spectrum learned by the model, while the red points show the rescaled exact eigenvalues ℓ(ℓ+1)/(j+w)2 with multiplicities 2ℓ+1 , where ℓ=0,1,…,2j=3 .
Figure 5: Left: Dataset of 2000 points sampled uniformly from the surfaces of two spheres centered at z=0 and z=3 with radii 1.5 and 1, respectively. The surfaces are shown in gray. Gaussian noise with standard deviation η=0.1 was added independently to each coordinate of each point. Right: Quantum geometric representation learned via trained quantum operators X . The QCML point cloud is shown, with point colors indicating the local Berry curvature. Seven degeneracy points of the displacement Hamiltonian (monopoles) in the illustrated run are highlighted in red and blue, corresponding to topological charges ±1 , respectively.
Figure 6: Left: Quantum geometry point cloud derived from the learned quantum geometry. Points are colored by the uncertainty σ(x) evaluated at each location. Seven degeneracy points (monopoles) of the displacement Hamiltonian in the illustrated run are highlighted in red and blue, corresponding to topological charges ±1 , respectively. Right: Partial spectrum of the matrix Laplacian derived from the learned quantum geometry. The first 12 eigenvalues are shown. A spectral gap separates the first two eigenvalues from the rest, consistent with the presence of two almost disconnected components in the data.
Figure 7: Left: Dataset of 2000 points sampled on the surface of a 2D unit sphere with probability density proportional to (1+cosθ)3 with respect to surface area. The unit sphere is shown in gray for reference. Gaussian noise with standard deviation η=0.1 was added independently to each coordinate of each point. Right: Quantum geometric representation learned via trained quantum operators X . The QCML point cloud is shown, with point colors indicating the local Berry curvature. Seven degeneracy points of the displacement Hamiltonian (monopoles) are highlighted in blue, corresponding to topological charges −1 .
Figure 8: Left: Quantum geometry point cloud derived from the learned quantum geometry. Points are colored by the uncertainty σ(x) evaluated at each location. Seven degeneracy points (monopoles) of the displacement Hamiltonian are highlighted in blue, corresponding to topological charge −1 . Right: Partial spectrum of the matrix Laplacian derived from the learned quantum geometry. The first 12 eigenvalues are shown. The clustering of eigenvalues into groups of 1, 3, and 5 reflects the characteristic spectrum of the Laplacian on a sphere, corresponding to angular momentum modes ℓ=0,1,2 , respectively.
Figure 9: Examples of conformal maps from the dataset. Each map is represented by the image (black points) of 100 reference points (shown in the leftmost panel) under a conformal transformation f(z;at,θ=0) , where at is sampled uniformly from the disk ∣a∣<0.9 . The corresponding QCML-learned quantum geometric representations are shown in orange. The associated complex parameters a are displayed in polar coordinates.
Figure 10: Left: Spectrum of the response matrix at all data points for the conformal map dataset. Gray lines show the spread of eigenvalues; black dots indicate the mean. The gap after the first two eigenvalues suggests intrinsic dimension d=2 . Right: Spectral counting function N(λ) vs eigenvalue λ of the matrix Laplacian. The approximate linear growth of N(λ)∼λd/2 is consistent with d=2 as an intrinsic dimension of the data.
Figure 11: Overlap matrix between the learned quantum observables Xa and the eigenmodes Yi of the matrix Laplacian. The size and intensity of circles indicate the magnitude of the overlap Tr(YiXa) . Only a subset of the eigenmodes is shown. The strong alignment of Xa with only first two leading eigenmodes Y1,2 suggests that these correspond to the dominant geometric directions in the feature space. Only a portion of the full overlap matrix is shown.
Figure 12: QCML cloud for the conformal map dataset visualized using the first two Laplacian eigenmaps (Y1,Y2) . Left: Points are colored by the modulus ∣a∣ of the complex parameter generating the map. Right: Points are colored by the phase arg(a) .
Figure 13: Quantum geometry of the conformal map dataset visualized via Laplacian eigenmaps (Y1,Y2,Y3) . Left: Points colored by modulus ∣a∣ of the conformal parameter. Right: Same embedding, colored by argument arg(a) .
w
0.02
0.05
0.10
0.20
0.50
0.80
d
7
5
4
3
2
1
Table 1: Estimated intrinsic dimension d as a function of the parameter w . The values of d are computed as the number of averaged response eigenvalues greater than the conventional threshold of 0.5 .
Figure 14: Left: Spectrum of the normalized matrix Laplacian Δ/30 derived from the learned quantum geometry at w=0.2 . The absence of degenerate zero modes indicates that the data geometry has no disconnected components. Right: Overlap matrix between the learned quantum observables Xa (rows), corresponding to 30 features of the Wisconsin Breast Cancer Data, and the eigenmodes Yi of the matrix Laplacian (columns). The size and intensity of circles indicate the magnitude of the overlap Tr(YiXa) . Only a subset of the eigenmodes is shown.
Representation
Classifier
ROC AUC
Accuracy
Eigenmaps, w=0.2 (3 modes)
Logistic regression
0.973±0.019
0.907±0.046
Eigenmaps, w=0.2 (3 modes)
Gradient boosting
0.981±0.014
0.938±0.024
Eigenmaps, w=0.5 (3 modes)
Logistic regression
0.964±0.022
0.912±0.038
Eigenmaps, w=0.5 (3 modes)
Gradient boosting
0.970±0.026
0.930±0.032
PCA (3 components)
Logistic regression
0.990±0.010
0.948±0.026
PCA (3 components)
Gradient boosting
0.985±0.015
0.942±0.025
Table 2: Cross-validated discrimination of benign from malignant samples. Values are means ± one standard deviation across ten stratified folds.
Figure 15: Quantum geometry of the WBC dataset at w=0.2 , visualized using Laplacian eigenmaps (Y1,Y2,Y3) . Left: points colored by diagnosis labels (red for malignant and green for benign samples). Right: projection onto the two leading eigenmaps (Y1,Y2) .
Figure 16: Left: Quantum geometry of the WBC dataset visualized using Laplacian eigenmaps (Y1,Y2,Y3) . The expectation values of Y1 collapse onto two distinct levels, placing 567 samples on a nearly two-dimensional surface and two nearly coincident malignant samples at one visible isolated location. Points are colored by diagnosis labels (red for malignant and green for benign samples). Right: Spectrum of the normalized matrix Laplacian Δ/30 derived from the learned quantum geometry. The lowest part of the spectrum contains three very small eigenvalues, indicating that the quantum geometry consists of three nearly disconnected components.
Many datasets encountered across a wide range of domains possess rich geometric and topological structure that is difficult to capture using conventional vector-based representations. Quantum machine learning offers the possibility of processing high-dimensional data in Hilbert spaces, but its practical success depends critically on how classical data is encoded into quantum states. We introduce quantum topological data encoding (QTDE), a general framework for encoding topological information into quantum states via topology-driven quantum evolution. Our method generalises an existing topology-driven quantum encoding framework to higher-dimensional data. We test the proposed method on clique-complexes classification tasks, and provide preliminary evidence that topology-driven quantum representations can capture discriminative information beyond that available through direct comparisons of classical topological descriptors. The proposed quantum representations consistently outperform a baseline based on direct comparisons of the combinatorial Laplacians describing the underlying topological structure. We indicate several areas of application where the framework can be used to provide a more efficient and reliable data representation.
Adam Wesołowski, Dimitrios Thanos, Daniel Leykam +1
Royal Holloway University of London, Department of Computer Science, UK · Leiden Institute of Advanced Computer Science (LIACS), Leiden University, Leiden, The Netherlands · Science, Mathematics and Technology Cluster, Singapore University of Technology and Design, Singapore +1
Quantum machine learning is often motivated by the idea that quantum systems can expose useful high-dimensional structure that is difficult to access with classical models. We isolate one central component of this claim: the fixed data-encoding map. Amplitude, angle, and basis encoding are evaluated as deterministic feature maps for classical supervised learning under matched output dimensionality and strong classical controls. The benchmark compares these encodings against raw linear models, random Fourier features, polynomial features, PCA, RBF SVMs, and shallow neural networks across diverse classical datasets. Rather than treating performance as a single endpoint, we analyze the geometry of each representation through effective rank, condition number, centered kernel alignment, predictive performance, and practical overhead. The resulting picture is mechanistic: amplitude encoding can remove magnitude information through unit-sphere normalization, angle encoding can become geometrically redundant with raw linear features, and basis encoding can impose a binary Hamming geometry that is poorly aligned with smooth decision structure. These findings do not argue against quantum computation, however, they show that fixed quantum-inspired encoding geometry alone is not a reliable source of machine-learning advantage on classical data.
Toheeb Ogunade, Taofeek Kassim, Etinosa Osaro
Department of Computer Science, University of Lagos, Lagos, Nigeria · Department of Physics, University of Lagos, Lagos, Nigeria · Department of Chemical Engineering, University of Notre Dame, Indiana, USA
Quantum Machine Learning (QML) has a theoretical advantage over classical algorithms for specific synthetic data that is computationally hard. However, it remains unclear whether such theoretical advantages play a meaningful role for realistic classical data, such as natural images. Quantum structural properties induced by unitary operations or entanglement introduce inductive biases that differ from those of classical models, potentially leading to distinct learning behavior on real-world data. In a fully quantum learning pipeline, classical real-world data are first encoded into quantum states via a quantum feature mapping, analogous to a classical feature extractor, and a quantum classifier then classifies the resulting states. Quantum feature mappings generate representations that are difficult to realize classically, while also inducing representation-level biases that do not necessarily match the structure of realistic data. To enable a fair comparison at the classifier level, we focus on the classifier and use classical feature extractors that already provide powerful representations for natural images. By controlling the data scale and supervision strength, we analyze how the unitary structure of quantum classifiers affects generalization, and show that their effectiveness depends strongly on the geometry of the input representations. These results indicate that quantum classifiers benefit from classical features that are aligned with quantum feature spaces. Our analysis shows that kernel alignment identifies representations compatible with quantum feature spaces, providing insights into translating classical models into quantum algorithms.
Yu Yamaoka, Weng Ian Chan, Kosuke Mitarai
Center for Quantum Information and Quantum Biology, The University of Osaka, 1-2 Machikaneyama, Toyonaka, 560-0043, Osaka, Japan. · SANKEN, The University of Osaka, 1-5 Yamadaoka, Suita, 565-0871, Osaka, Japan. · Graduate School of Engineering Science, The University of Osaka, 1-3 Machikaneyama, Toyonaka, 560-8531, Osaka, Japan.