Abstract:A seminal open question of Pisier and Mendel--Naor asks whether every degree-regular graph which satisfies the classical discrete Poincaré inequality for scalar functions, also satisfies an analogous inequality for functions taking values in \textit{any} normed space with non-trivial cotype. Motivated by applications, it is also greatly important to quantify the dependence of the corresponding optimal Poincaré constant on the cotype $q$. Works of Odell--Schlumprecht (1994), Ozawa (2004), and Naor (2014) make substantial progress on the former question by providing a positive answer for normed spaces which also have an unconditional basis, in addition to finite cotype. However, little is known in the way of quantitative estimates: the mentioned results imply a bound on the Poincaré constant depending super-exponentially on $q$. We introduce a novel combinatorial framework for proving quantitative nonlinear spectral gap estimates. The centerpiece is a property of regular graphs that we call \emph{long range expansion}, which holds with high probability for random regular graphs. Our main result is that any regular graph with the long-range expansion property satisfies a discrete Poincaré inequality for any normed space with an unconditional basis and cotype $q$, with a Poincaré constant that depends \emph{polynomially} on $q$, which is optimal. As an application, any normed space with an unconditional basis which admits a low distortion embedding of an $n$-vertex random regular graph, must have cotype at least polylogarithmic in $n$. This extends a celebrated lower-bound of Matoušek for low distortion embeddings of random graphs into $\ell_q$ spaces.

Spectral Norm of Convolutional Layers with Circular and Zero Paddings

Tight and Efficient Upper Bound on Spectral Norm of Convolutional Layers

Spectral Representations for Convolutional Neural Networks

Depthwise Separable Convolutions Allow for Fast and Memory-Efficient Spectral Normalization

Spectral bounding: Strictly satisfying the 1-Lipschitz property for generative adversarial networks

Spectrality of a class of infinite convolutions with and without compact supports in $\mathbb{R}^d$

Dilated Convolution with Learnable Spacings: beyond bilinear interpolation

A Frobenius Norm Regularization Method for Convolutional Kernel Tensors in Neural Networks

Automorphic Spectra and the Conformal Bootstrap

A combinatorial approach to nonlinear spectral gaps

The Singular Values of Convolutional Layers

Reduce Computational Complexity for Convolutional Layers by Skipping Zeros

Padding-free Convolution based on Preservation of Differential Characteristics of Kernels

Breaking the Curse of Dimensionality with Convex Neural Networks

QuadConv: Quadrature-Based Convolutions with Applications to Non-Uniform PDE Data Compression

A Spectral Condition for Feature Learning

A Frobenius norm regularization method for convolutional kernels to avoid unstable gradient problem

The Sample Complexity of One-Hidden-Layer Neural Networks

Spectrum Extraction and Clipping for Implicitly Linear Layers

Spectral GNN via Two-dimensional (2-D) Graph Convolution

Spectral Leakage and Rethinking the Kernel Size in CNNs