Larry Goldstein

Krishnakumar Balasubramanian, Larry Goldstein, Nathan Ross et al.

We derive upper bounds on the Wasserstein distance ($W_1$), with respect to $\sup$-norm, between any continuous $\mathbb{R}^d$ valued random field indexed by the $n$-sphere and the Gaussian, based on Stein's method. We develop a novel Gaussian smoothing technique that allows us to transfer a bound in a smoother metric to the $W_1$ distance. The smoothing is based on covariance functions constructed using powers of Laplacian operators, designed so that the associated Gaussian process has a tractable Cameron-Martin or Reproducing Kernel Hilbert Space. This feature enables us to move beyond one dimensional interval-based index sets that were previously considered in the literature. Specializing our general result, we obtain the first bounds on the Gaussian random field approximation of wide random neural networks of any depth and Lipschitz activation functions at the random field level. Our bounds are explicitly expressed in terms of the widths of the network and moments of the random weights. We also obtain tighter bounds when the activation function has three bounded derivatives.

Larry Goldstein

Given a non-negative random variable $W$ and $θ>0$, let the generalized Dickman transformation map the distribution of $W$ to that of $$ W^*=_d U^{1/θ}(W+1), $$ where $U \sim {\cal U}[0,1]$, a uniformly distributed variable on the unit interval, independent of $W$, and where $=_d$ denotes equality in distribution. It is well known that $W^*$ and $W$ are equal in distribution if and only if $W$ has the generalized Dickman distribution ${\cal D}_θ$. We demonstrate that the Wasserstein distance $d_1$ between $W$, a non-negative random variable with finite mean, and $D_θ$ having distribution ${\cal D}_θ$ obeys the inequality $$ d_1(W,D_θ) \le (1+θ)d_1(W,W^*). $$ The specialization of this bound to the case $θ=1$ and coupling constructions yield $$ d_1(W_{n,1},D) \le \frac{8\log (n/2)+10}{n} \quad \mbox{for all $n \ge 1$, where} \quad W_{n,1}=\frac{1}{n}C_{n,1}-1, $$ and $C_{n,m}$ is the number of comparisons made by the Quickselect algorithm to find the $m^{th}$ smallest element of a list of $n$ distinct numbers. A similar bound holds for $m \ge 2$, and together recover the results of [12] that show distributional convergence of $W_n$ to the standard Dickman distribution in the asymptotic regime $m=o(n)$. By developing an exact expression for the expected running time $E[C_{n,m}]$, lower bounds are provided that show the rate is not improvable for all $m \not = 2$.