Abstract:It is well known that fluctuations of marginals of high-dimensional random vectors that satisfy a certain concentration estimate called the thin shell condition are approximately Gaussian. In this article we identify a general condition on a sequence of high-dimensional random vectors under which one can identify the exponential decay rate of large deviation probabilities of the corresponding sequence of marginals. More precisely, consider the projection of an n-dimensional random vector onto a random k n -dimensional basis, k n ≤ n , drawn uniformly from the Haar measure on the Stiefel manifold of orthonormal k n -frames in R n , in three different asymptotic regimes as n → ∞ : “constant” ( k n = k ), “sublinear” ( k n → ∞ but k n / n → 0 ) and “linear” ( k n / n → λ with 0 < λ ≤ 1 ). When the sequence of random vectors satisfies a certain “asymptotic thin shell condition”, we establish large deviation principles for the corresponding sequence of random projections in the constant regime, and for the sequence of empirical measures of the coordinates of the random projections in the sublinear and linear regimes. We also establish large deviation principles for scaled ℓ q norms of the random projections in all three regimes. Moreover, we show that the asymptotic thin shell condition holds for various sequences of random vectors of interest, including the uniform measure on suitably scaled ℓ p n balls, for p ∈ [ 1 , ∞ ) , and generalized Orlicz balls defined via a superquadratic function, as well as a class of Gibbs measures with superquadratic interaction potential. Along the way, we obtain logarithmic asymptotics of volumes of high-dimensional Orlicz balls, which may be of independent interest. We also show that the decay rate of large deviation probabilities of Euclidean norms of multi-dimensional projections of ℓ p n balls, when p ∈ [ 1 , 2 ) , exhibits an unexpected phase transition in the sublinear regime, thus disproving an earlier conjecture due to Alonso-Gutiérrez et al. Random projections of high-dimensional random vectors are of interest in a range of fields including asymptotic convex geometry and high-dimensional statistics.

Optional and predictable projections of normal integrands and convex-valued processes

Projection and Convolution Operations for Integrally Convex Functions

And$\hat{o}$-Douglas type characterization of generalized conditional expectations, optional projections and predictable projections

O(log T) Projections for Stochastic Optimization of Smooth and Strongly Convex Functions

Strong Duality in Risk-Constrained Nonconvex Functional Programming

On the differentiability of $ϕ$-projections in the discrete finite case

Dual Solutions in Convex Stochastic Optimization

A General Theory of Operator-Valued Measures

Convex duality for stochastic singular control problems

Isospin dependence of nucleon effective mass in Dirac Brueckner-Hartree-Fock approach

Projection Theorems in the Presence of Expansions

Probability Tools for Sequential Random Projection

Covering Constants for Metric Projection Operator with Applications to Stochastic Fixed-Point Problems

Stochastic Proximal Point Methods for Monotone Inclusions under Expected Similarity

A Fenchel-Moreau-Rockafellar type theorem on the Kantorovich-Wasserstein space with Applications in Partially Observable Markov Decision Processes

An asymptotic thin shell condition and large deviations for random multidimensional projections

Set-valued stochastic integrals for convoluted Lévy processes

Random projections beyond zero overlap

On semi-continuity and continuity of the smallest and largest minimizing point of real convex functions with applications in probability and statistics

Unsafe Probabilities and Risk Contours for Stochastic Processes using Convex Optimization

An extension of Minkowski's theorem and its applications to questions about projections for measures