arXiv:2610.06080
Gaussian Universality and Its Breakdown in Tensor-Network Machine Learning
Shi-Tuan Wang, Zidu Liu, Li-Wei Yu
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.
AI Summary
The paper asks when randomly initialized tensor-network outputs admit Gaussian-process descriptions, and whether increasing bond dimension and physical dimension have equivalent effects. It studies expectation values in periodic MPS constructed from independent Haar-random unitary embeddings, comparing single-site observables with fully supported Pauli strings and specified linear combinations. For fixed physical dimension $d$ and growing bond dimension $D$ and system size $n$, variance-normalized outputs approach joint Gaussian distributions under the stated scaling conditions. For Pauli-normalized local observables, the variance scales as $1/(D^2d)$ when $D^2\ll d^n$, and as $d^{-n}$ when $D^2\gg d^n$. Global outputs have variance $d^{-n}$; their Gaussian limit requires $D/n\to\infty$, with relative moment corrections bounded by $O(nk!m^{k/2}/D)+O(k^2/d^n)$ for moment order $k$ and fixed basis size $m$. At fixed $(D,n\ge2)$ and $d\to\infty$, local outputs within the specified Clifford span remain Gaussian, with moment corrections $O(k^2D^2/d)$. Fully supported Pauli outputs remain non-Gaussian: their limiting excess kurtosis is $(1+2/D^2)^n+2(1-1/D^2)^n-3$. At $D=1$, their rescaled limit is a product of $n$ independent Gaussians. The analysis uses Haar integration, permutation-indexed transfer matrices, moment convergence, and Cramér–Wold arguments, supported by sampling. Scope is restricted to this random MPS ensemble and specified asymptotic regimes; global joint-Gaussian results impose Clifford-algebra restrictions. Explicit data encodings, prediction kernels, noisy inference, intermediate-support observables, and other network geometries are left for future work.
Central claim may not hold: The main-text global GP theorem omits a relative-growth condition on bond dimension and system size, and its stated covariance fails along d=2, n=D³ even though both D and n diverge. The abstract's large-physical-dimension Gaussian claim for local observables requires spectral restrictions: at bond dimension one, a centered single-site rank-one projector has asymptotically exponential rather than Gaussian rescaled fluctuations.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 8+42pages, comments are welcome