Infinite-Width Limit of a Single Attention Layer: Analysis via Tensor Programs

• URL: http://arxiv.org/abs/2506.00846v2
• Date: Sun, 26 Oct 2025 06:38:29 GMT
• Title: Infinite-Width Limit of a Single Attention Layer: Analysis via Tensor Programs
• Abstract summary: In this paper, we identify the infinite-width limit distribution of variables within a single attention layer under realistic architectural dimensionality.<n>We derive the exact form of this limit law without resorting to infinite-head approximations or tailored scalings.<n>Our findings lay the groundwork for developing a unified theory of deep Transformer architectures in the infinite-width regime.
• Score: 15.559242245579382
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: In modern theoretical analyses of neural networks, the infinite-width limit is often invoked to justify Gaussian approximations of neuron preactivations (e.g., via neural network Gaussian processes or Tensor Programs). However, these Gaussian-based asymptotic theories have so far been unable to capture the behavior of attention layers, except under special regimes such as infinitely many heads or tailored scaling schemes. In this paper, leveraging the Tensor Programs framework, we rigorously identify the infinite-width limit distribution of variables within a single attention layer under realistic architectural dimensionality and standard $1/\sqrt{n}$-scaling with $n$ dimensionality. We derive the exact form of this limit law without resorting to infinite-head approximations or tailored scalings, demonstrating that it departs fundamentally from Gaussianity. This limiting distribution exhibits non-Gaussianity from a hierarchical structure, being Gaussian conditional on the random similarity scores. Numerical experiments validate our theoretical predictions, confirming the effectiveness of our theory at finite width and accurate description of finite-head attentions. Beyond characterizing a standalone attention layer, our findings lay the groundwork for developing a unified theory of deep Transformer architectures in the infinite-width regime.

Related papers

• How Deep Are Deep GPs, Really? A Sharp Threshold and a Non-Gaussian Limit for Compositional GPs [48.096969031315744]
  Previous work has established that for the RBF kernel and a certain range of bandwidths $r$, the prior degenerates in the limit.<n>We prove that for $r$ below the threshold $r_c(d)$ the prior converges to a limit distribution $_barZ$.
  arXiv  Detail & Related papers  (2026-06-06T15:12:43Z)
• Pruning at Initialisation through the lens of Graphon Limit: Convergence, Expressivity, and Generalisation [16.628325681877556]
  Pruning at Initialisation methods discover sparse, trainableworks before training, but their theoretical mechanisms remain elusive.<n>In this work, we connect discrete intrinsic prunings to graph limit theory via graphons, establishing the graphon limit of PaI masks.<n>We derive two fundamental theoretical results: (i) a Universal Approximation Theorem for sparse networks that depends only on the dimension of active coordinate subspaces; and (ii) a Graphon-NTK generalisation bound demonstrating how the limit graphon geometry to align with informative features.
  arXiv  Detail & Related papers  (2026-02-06T13:02:47Z)
• Random-Matrix-Induced Simplicity Bias in Over-parameterized Variational Quantum Circuits [72.0643009153473]
  We show that expressive variational ansatze enter a Haar-like universality class in which both observable expectation values and parameter gradients concentrate exponentially with system size.<n>As a consequence, the hypothesis class induced by such circuits collapses with high probability to a narrow family of near-constant functions.<n>We further show that this collapse is not unavoidable: tensor-structured VQCs, including tensor-network-based and tensor-hypernetwork parameterizations, lie outside the Haar-like universality class.
  arXiv  Detail & Related papers  (2026-01-05T08:04:33Z)
• Precise gradient descent training dynamics for finite-width multi-layer neural networks [8.057006406834466]
  We provide the first precise distributional characterization of gradient descent iterates for general multi-layer neural networks.<n>Our non-asymptotic state evolution theory captures Gaussian fluctuations in first-layer weights and concentration in deeper-layer weights.
  arXiv  Detail & Related papers  (2025-05-08T02:19:39Z)
• A Near Complete Nonasymptotic Generalization Theory For Multilayer Neural Networks: Beyond the Bias-Variance Tradeoff [57.25901375384457]
  We propose a nonasymptotic generalization theory for multilayer neural networks with arbitrary Lipschitz activations and general Lipschitz loss functions.<n>In particular, it doens't require the boundness of loss function, as commonly assumed in the literature.<n>We show the near minimax optimality of our theory for multilayer ReLU networks for regression problems.
  arXiv  Detail & Related papers  (2025-03-03T23:34:12Z)
• Proportional infinite-width infinite-depth limit for deep linear neural networks [0.16385815610837165]
  We study the distributional properties of linear neural networks with random parameters in the context of large networks, where the number of layers diverges in proportion to the number of neurons per layer. We explore the joint proportional limit in which both depth and width diverge but maintain a constant ratio, yielding a non-Gaussian distribution that retains correlations between outputs.
  arXiv  Detail & Related papers  (2024-11-22T11:25:52Z)
• Generalization of Geometric Graph Neural Networks with Lipschitz Loss Functions [84.01980526069075]
  We study the generalization capabilities of geometric graph neural networks (GNNs)<n>We prove a generalization gap between the optimal empirical risk and the optimal statistical risk of this GNN.<n>We verify this theoretical result with experiments on multiple real-world datasets.
  arXiv  Detail & Related papers  (2024-09-08T18:55:57Z)
• Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion [83.90492831583997]
  We show that a batch-normalized network can keep the optimal signal propagation properties, but avoid exploding gradients in depth. We use a Multi-Layer Perceptron (MLP) with linear activations and batch-normalization that provably has bounded depth. We also design an activation shaping scheme that empirically achieves the same properties for certain non-linear activations.
  arXiv  Detail & Related papers  (2023-10-03T12:35:02Z)
• Posterior Inference on Shallow Infinitely Wide Bayesian Neural Networks under Weights with Unbounded Variance [1.5960546024967326]
  It is known that the infinite width scaling limit of a Bayesian neural network with one hidden layer is a Gaussian process, when the network weights have bounded prior variance. Neal's result has been extended to networks with multiple hidden layers and to convolutional neural networks, also with Gaussian process scaling limits. Our contribution is an interpretable and computationally efficient procedure for posterior inference, using a conditionally Gaussian representation, that then allows full use of the Gaussian process machinery for tractable posterior inference and uncertainty quantification in the non-Gaussian regime.
  arXiv  Detail & Related papers  (2023-05-18T02:55:00Z)
• A Unified Algebraic Perspective on Lipschitz Neural Networks [88.14073994459586]
  This paper introduces a novel perspective unifying various types of 1-Lipschitz neural networks. We show that many existing techniques can be derived and generalized via finding analytical solutions of a common semidefinite programming (SDP) condition. Our approach, called SDP-based Lipschitz Layers (SLL), allows us to design non-trivial yet efficient generalization of convex potential layers.
  arXiv  Detail & Related papers  (2023-03-06T14:31:09Z)
• Unified Field Theory for Deep and Recurrent Neural Networks [56.735884560668985]
  We present a unified and systematic derivation of the mean-field theory for both recurrent and deep networks. We find that convergence towards the mean-field theory is typically slower for recurrent networks than for deep networks. Our method exposes that Gaussian processes are but the lowest order of a systematic expansion in $1/n$.
  arXiv  Detail & Related papers  (2021-12-10T15:06:11Z)
• The Future is Log-Gaussian: ResNets and Their Infinite-Depth-and-Width Limit at Initialization [18.613475245655806]
  We study ReLU ResNets in the infinite-depth-and-width limit, where both depth and width tend to infinity as their ratio, $d/n$, remains constant. Using Monte Carlo simulations, we demonstrate that even basic properties of standard ResNet architectures are poorly captured by the Gaussian limit.
  arXiv  Detail & Related papers  (2021-06-07T23:47:37Z)
• Neural Networks and Quantum Field Theory [0.0]
  We propose a theoretical understanding of neural networks in terms of Wilsonian effective field theory. The correspondence relies on the fact that many neural networks are drawn from Gaussian processes.
  arXiv  Detail & Related papers  (2020-08-19T18:00:06Z)
• Generalization bound of globally optimal non-convex neural network training: Transportation map estimation by infinite dimensional Langevin dynamics [50.83356836818667]
  We introduce a new theoretical framework to analyze deep learning optimization with connection to its generalization error. Existing frameworks such as mean field theory and neural tangent kernel theory for neural network optimization analysis typically require taking limit of infinite width of the network to show its global convergence.
  arXiv  Detail & Related papers  (2020-07-11T18:19:50Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.