Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.06080

Gaussian Universality and Its Breakdown in Tensor-Network Machine Learning

Shi-Tuan Wang, Zidu Liu, Li-Wei Yu

quant-phcs.LGSubmitted Oct 5, 2026• Announced Oct 6, 2026• 0 citations
arXivHTMLPDF

Abstract

Gaussian-process limits are powerful in describing overparameterized machine learning models, yet their validity in structured tensor-network architectures remains unclear. Here we analytically present a moment-based approach that identifies precise conditions for the emergence and breakdown of Gaussian universality in tensor-network learning models, with a focus on matrix product states. We prove that in the large bond dimension limit, the learning models with both local and global observables converge to Gaussian processes, with explicit finite-size bounds on higher-order moment deviations. Whereas in the large physical dimension limit, the Gaussian universality no longer persists: while the models with local observables retain Gaussian-process behavior, those global cases exhibit persistent non-Gaussian corrections. Our results reveal that Gaussian-process behavior in tensor-network learning is controlled not only by parameter number, but also by architectural scaling, observable locality, and the spectral properties.

Comments
8+42pages, comments are welcome

Related papers

Presence and Absence of Barren Plateaus in Tensor-Network Based Machine Learning

Zidu Liu, Li-Wei Yu, Luming Duan +1 more

Tensor networks are efficient representations of high-dimensional tensors with widespread applications in quantum many-body physics. Recently, they have been adapted to the field of machine learning, giving rise to an emergent research frontier that has attracted considerable attention. Here, we study the trainability of tensor-network based machine learning models by exploring the landscapes of different loss functions, with a focus on the matrix product states (also called tensor trains) architecture. In particular, we rigorously prove that barren plateaus (i.e., exponentially vanishing gradients) prevail in the training process of the machine learning algorithms with global loss functions. Whereas, for local loss functions the gradients with respect to variational parameters near the local observables do not vanish as the system size increases. Therefore, the barren plateaus are absent in this case and the corresponding models could be efficiently trainable. Our results reveal a crucial aspect of tensor-network based machine learning in a rigorous fashion, which provide a valuable guide for both practical applications and theoretical studies in the future.

arXiv:2108.08312•Physical Review Letters•Dec 28, 2022•49 citations

Random Gaussian Augmented Matrix Product States

Stavya Puri, Poetri Sonya Tarabunga, Piotr Sierant

Strong entanglement limits the reach of tensor-network descriptions of quantum many-body systems. We study Gaussian-augmented matrix product states (GAMPS), formed by applying a free-fermionic (Gaussian) unitary to a matrix product state. To explore the capabilities of this family, we consider random GAMPS}, a minimally structured ensemble obtained by independently sampling random matrix product states and Gaussian unitaries. Exploiting the algebraic structure of Gaussian unitaries, we develop a replica tensor-network method that computes averages over the random GAMPS ensemble for hundreds of qubits. Random GAMPS exhibit nearly maximal volume-law entanglement already at modest bond dimension, while other quantum resource quantifiers rapidly approach Haar-random state values as the bond dimension increases. The entanglement scaling suggests an exponential reduction in the bond dimension needed to approach the volume-law entropy compared with unaugmented matrix product states. We find a state two-design behavior of random GAMPS at fixed accuracy with a bond dimension independent of system size. Our results reveal how Gaussian augmentation expands the expressive power of tensor networks.

quant-pharXiv:2609.25458•Sep 23, 2026•0 citations

Universal Scaling Laws in Quantum-Probabilistic Machine Learning by Tensor Network: Toward Interpreting Representation and Generalization Powers

Sheng-Chen Bai, Shi-Ju Ran

Abstract The interpretation of representations and generalization powers has been a long-standing challenge in the fields of machine learning (ML) and artificial intelligence. This study contributes to understanding the emergence of universal scaling laws in quantum-probabilistic ML. We consider the generative tensor network (GTN) in the form of a matrix-product state as an example and show that with an untrained GTN (such as a random TN state), the negative logarithmic likelihood (NLL) L generally increases linearly with the number of features M, that is, L ≃ kM + const. This is a consequence of the so-called “catastrophe of orthogonality,” which states that quantum many-body states tend to become exponentially orthogonal to each other as M increases. This study reveals that, while gaining information through training, the linear-scaling law is suppressed by a negative quadratic correction, leading to L ≃ βM – αM 2 + const. The scaling coefficients exhibit logarithmic relationships with the number of training samples and quantum channels χ. The emergence of a quadratic correction term in the NLL for the testing (training) set can be regarded as evidence of the generalization (representation) power of the GTN. Over-parameterization can be identified by the deviation in the values of α between the training and testing sets while increasing χ. We further investigate how orthogonality in the quantum-feature map relates to the satisfaction of quantum-probabilistic interpretation and the representation and generalization powers of the GTN. Unveiling universal scaling laws in quantum-probabilistic ML would be a valuable step toward establishing a white-box ML scheme interpreted within the quantum-probabilistic framework.

arXiv:2410.09703•Chinese Physics Letters•Nov 19, 2024•3 citations

Algorithmic Locality of Quantum Tensor Network Belief Propagation

Siddhant Midha, Yifan F. Zhang, Daniel Malz +2 more

Belief propagation has recently emerged as a powerful framework for evaluating tensor networks in higher dimensions, combining computational efficiency with provable analytical guarantees. In this work, we develop the first end-to-end theory of tensor network belief propagation for a class of projected entangled pair states satisfying \emph{strong injectivity}. We show that when the injectivity parameter exceeds a constant threshold, BP fixed points can be found efficiently, and a cluster-corrected BP algorithm computes physical quantities to $1/\mathrm{poly}(N)$ error in $\mathrm{poly}(N)$ time for an $N$ qubit system. We further identify a striking phenomenon we term \emph{algorithmic locality}: local perturbations of the tensor network affect the BP fixed point with an influence decaying rapidly with distance. Moreover, through the cluster expansion, this locality extends to observables, implying that local expectation values can be approximated from local data with controlled accuracy. We exhibit three regimes admitting strong injectivity: random Gaussian PEPS with large physical dimension, high-temperature Gibbs states of commuting two-body Hamiltonians, and fixed-depth quantum circuits subject to sufficiently strong depolarizing noise. Our bounds are worst-case sufficient conditions and do not identify sharp algorithmic transitions.

quant-pharXiv:2604.21919•Apr 24, 2026•0 citations

Lower and Upper Bounds on the VC-Dimension of Tensor Network Models

Behnoush Khavari, Guillaume Rabusseau

Tensor network methods have been a key ingredient of advances in condensed matter physics and have recently sparked interest in the machine learning community for their ability to compactly represent very high-dimensional objects. Tensor network methods can for example be used to efficiently learn linear models in exponentially large feature spaces [Stoudenmire and Schwab, 2016]. In this work, we derive upper and lower bounds on the VC dimension and pseudo-dimension of a large class of tensor network models for classification, regression and completion. Our upper bounds hold for linear models parameterized by arbitrary tensor network structures, and we derive lower bounds for common tensor decomposition models~(CP, Tensor Train, Tensor Ring and Tucker) showing the tightness of our general upper bound. These results are used to derive a generalization bound which can be applied to classification with low rank matrices as well as linear classifiers based on any of the commonly used tensor decomposition models. As a corollary of our results, we obtain a bound on the VC dimension of the matrix product state classifier introduced in [Stoudenmire and Schwab, 2016] as a function of the so-called bond dimension~(i.e. tensor train rank), which answers an open problem listed by Cirac, Garre-Rubio and Pérez-García in [Cirac et al., 2019].

arXiv:2106.11827•Jun 22, 2021•3 citations

Tensor network finite-size scaling for two-dimensional 3-state clock model

Debasmita Maiti, Sing-Hong Chan, Pochung Chen

Abstract We benchmark recently proposed tensor network based finite-size scaling analysis in (Phys. Rev. B 2023 107 205123) against two-dimensional classical 3-state clock model. Due to the higher complexity of the model, more complicated crossover behavior is observed. We advocate that the crossover behavior can be understood from the perspective of finite bond dimension inducing relevant perturbation. This leads to a general strategy to best estimate the critical properties for a given set of control parameters. For the critical temperature T c , the relative error at the order of 10−7 can be reached with bond dimension D = 70. On the other hand, with bond dimension D = 60, the relative errors of the critical exponents ν , β , α are at the order of 10−2. Increasing the bond dimension to D = 90, these relative errors can be reduced at least to the order of 10−3. In all cases our results indicate that the errors can be systematically reduced by increasing the bond dimension and the stacking number.

arXiv:2312.14002•New Journal of Physics•Apr 24, 2025•1 citation
Gaussian-process limits are powerful in describing overparameterized machine learning models, yet their validity in structured tensor-network architectures remains unclear. Here we analytically present a moment-based approach that identifies precise conditions for the emergence and breakdown of Gaussian universality in tensor-network learning models, with a focus on matrix product states. We prove that in the large bond dimension limit, the learning models with both local and global observables converge to Gaussian processes, with explicit finite-size bounds on higher-order moment deviations. Whereas in the large physical dimension limit, the Gaussian universality no longer persists: while the models with local observables retain Gaussian-process behavior, those global cases exhibit persistent non-Gaussian corrections. Our results reveal that Gaussian-process behavior in tensor-network learning is controlled not only by parameter number, but also by architectural scaling, observable locality, and the spectral properties.
Tensor networks are efficient representations of high-dimensional tensors with widespread applications in quantum many-body physics. Recently, they have been adapted to the field of machine learning, giving rise to an emergent research frontier that has attracted considerable attention. Here, we study the trainability of tensor-network based machine learning models by exploring the landscapes of different loss functions, with a focus on the matrix product states (also called tensor trains) architecture. In particular, we rigorously prove that barren plateaus (i.e., exponentially vanishing gradients) prevail in the training process of the machine learning algorithms with global loss functions. Whereas, for local loss functions the gradients with respect to variational parameters near the local observables do not vanish as the system size increases. Therefore, the barren plateaus are absent in this case and the corresponding models could be efficiently trainable. Our results reveal a crucial aspect of tensor-network based machine learning in a rigorous fashion, which provide a valuable guide for both practical applications and theoretical studies in the future.
Strong entanglement limits the reach of tensor-network descriptions of quantum many-body systems. We study Gaussian-augmented matrix product states (GAMPS), formed by applying a free-fermionic (Gaussian) unitary to a matrix product state. To explore the capabilities of this family, we consider random GAMPS}, a minimally structured ensemble obtained by independently sampling random matrix product states and Gaussian unitaries. Exploiting the algebraic structure of Gaussian unitaries, we develop a replica tensor-network method that computes averages over the random GAMPS ensemble for hundreds of qubits. Random GAMPS exhibit nearly maximal volume-law entanglement already at modest bond dimension, while other quantum resource quantifiers rapidly approach Haar-random state values as the bond dimension increases. The entanglement scaling suggests an exponential reduction in the bond dimension needed to approach the volume-law entropy compared with unaugmented matrix product states. We find a state two-design behavior of random GAMPS at fixed accuracy with a bond dimension independent of system size. Our results reveal how Gaussian augmentation expands the expressive power of tensor networks.
Abstract The interpretation of representations and generalization powers has been a long-standing challenge in the fields of machine learning (ML) and artificial intelligence. This study contributes to understanding the emergence of universal scaling laws in quantum-probabilistic ML. We consider the generative tensor network (GTN) in the form of a matrix-product state as an example and show that with an untrained GTN (such as a random TN state), the negative logarithmic likelihood (NLL) L generally increases linearly with the number of features M, that is, L ≃ kM + const. This is a consequence of the so-called “catastrophe of orthogonality,” which states that quantum many-body states tend to become exponentially orthogonal to each other as M increases. This study reveals that, while gaining information through training, the linear-scaling law is suppressed by a negative quadratic correction, leading to L ≃ βM – αM 2 + const. The scaling coefficients exhibit logarithmic relationships with the number of training samples and quantum channels χ. The emergence of a quadratic correction term in the NLL for the testing (training) set can be regarded as evidence of the generalization (representation) power of the GTN. Over-parameterization can be identified by the deviation in the values of α between the training and testing sets while increasing χ. We further investigate how orthogonality in the quantum-feature map relates to the satisfaction of quantum-probabilistic interpretation and the representation and generalization powers of the GTN. Unveiling universal scaling laws in quantum-probabilistic ML would be a valuable step toward establishing a white-box ML scheme interpreted within the quantum-probabilistic framework.
Belief propagation has recently emerged as a powerful framework for evaluating tensor networks in higher dimensions, combining computational efficiency with provable analytical guarantees. In this work, we develop the first end-to-end theory of tensor network belief propagation for a class of projected entangled pair states satisfying \emph{strong injectivity}. We show that when the injectivity parameter exceeds a constant threshold, BP fixed points can be found efficiently, and a cluster-corrected BP algorithm computes physical quantities to 1/poly(N)1/\mathrm{poly}(N)1/poly(N) error in poly(N)\mathrm{poly}(N)poly(N) time for an NNN qubit system. We further identify a striking phenomenon we term \emph{algorithmic locality}: local perturbations of the tensor network affect the BP fixed point with an influence decaying rapidly with distance. Moreover, through the cluster expansion, this locality extends to observables, implying that local expectation values can be approximated from local data with controlled accuracy. We exhibit three regimes admitting strong injectivity: random Gaussian PEPS with large physical dimension, high-temperature Gibbs states of commuting two-body Hamiltonians, and fixed-depth quantum circuits subject to sufficiently strong depolarizing noise. Our bounds are worst-case sufficient conditions and do not identify sharp algorithmic transitions.
Tensor network methods have been a key ingredient of advances in condensed matter physics and have recently sparked interest in the machine learning community for their ability to compactly represent very high-dimensional objects. Tensor network methods can for example be used to efficiently learn linear models in exponentially large feature spaces [Stoudenmire and Schwab, 2016]. In this work, we derive upper and lower bounds on the VC dimension and pseudo-dimension of a large class of tensor network models for classification, regression and completion. Our upper bounds hold for linear models parameterized by arbitrary tensor network structures, and we derive lower bounds for common tensor decomposition models~(CP, Tensor Train, Tensor Ring and Tucker) showing the tightness of our general upper bound. These results are used to derive a generalization bound which can be applied to classification with low rank matrices as well as linear classifiers based on any of the commonly used tensor decomposition models. As a corollary of our results, we obtain a bound on the VC dimension of the matrix product state classifier introduced in [Stoudenmire and Schwab, 2016] as a function of the so-called bond dimension~(i.e. tensor train rank), which answers an open problem listed by Cirac, Garre-Rubio and Pérez-García in [Cirac et al., 2019].
Abstract We benchmark recently proposed tensor network based finite-size scaling analysis in (Phys. Rev. B 2023 107 205123) against two-dimensional classical 3-state clock model. Due to the higher complexity of the model, more complicated crossover behavior is observed. We advocate that the crossover behavior can be understood from the perspective of finite bond dimension inducing relevant perturbation. This leads to a general strategy to best estimate the critical properties for a given set of control parameters. For the critical temperature T c , the relative error at the order of 10−7 can be reached with bond dimension D = 70. On the other hand, with bond dimension D = 60, the relative errors of the critical exponents ν , β , α are at the order of 10−2. Increasing the bond dimension to D = 90, these relative errors can be reduced at least to the order of 10−3. In all cases our results indicate that the errors can be systematically reduced by increasing the bond dimension and the stacking number.