arXiv subjects
Publications and source records attributed to Alexander M. Powell.
It is well-known that for a Gabor frame in L2, the canonical dual window is the dual window with the minimal L2-norm. In this paper, we address the problem of finding dual windows for a Gabor frame that minimizes the L1-norm. Since it has been shown that minimizing the L1-norm often coincides with minimizing the L0-norm, we also investigate the support of the dual Gabor window that achieves the minimal L1-norm.
The massive size of modern neural networks has motivated substantial recent interest in neural network quantization. We introduce Stochastic Markov Gradient Descent (SMGD), a discrete optimization method applicable to training quantized neural networks. The SMGD algorithm is designed for settings where memory is highly constrained during training. We provide theoretical guarantees of algorithm performance as well as encouraging numerical results.
We construct high order low-bit Sigma-Delta $(ΣΔ)$ quantizers for the vector-valued setting of fusion frames. We prove that these $ΣΔ$ quantizers can be stably implemented to quantize fusion frame measurements on subspaces $W_n$ using $\log_2( {\rm dim}(W_n)+1)$ bits per measurement. Signal reconstruction is performed using a version of Sobolev duals for fusion frames, and numerical experiments are given to validate the overall performance.
We prove that $L_2(\mathbb{R})$ contains a Schauder basis of non-negative functions. Similarly, $L_p(\mathbb{R})$ contains a Schauder basic sequence of non-negative functions such that $L_p(\mathbb{R})$ embeds into the closed span of the sequence. We prove as well that if $X$ is a separable Banach space with the bounded approximation property, then any set in $X$ with dense span contains a quasi-basis (Schauder frame) for $X$. Furthermore, if $X$ is a separable Banach lattice with a bibasis then any set in $X$ with dense span contains a u-frame.
A sharp version of the Balian-Low theorem is proven for the generators of finitely generated shift-invariant spaces. If generators $\{f_k\}_{k=1}^K \subset L^2(\mathbb{R}^d)$ are translated along a lattice to form a frame or Riesz basis for a shift-invariant space $V$, and if $V$ has extra invariance by a suitable finer lattice, then one of the generators $f_k$ must satisfy $\int_{\mathbb{R}^d} |x| |f_k(x)|^2 dx = \infty$, namely, $\widehat{f_k} \notin H^{1/2}(\mathbb{R}^d)$. Similar results are proven for frames of translates that are not Riesz bases without the assumption of extra lattice invariance. The best previously existing results in the literature give a notably weaker conclusion using the Sobolev space $H^{d/2+ε}(\mathbb{R}^d)$; our results provide an absolutely sharp improvement with $H^{1/2}(\mathbb{R}^d)$. Our results are sharp in the sense that $H^{1/2}(\mathbb{R}^d)$ cannot be replaced by $H^s(\mathbb{R}^d)$ for any $s<1/2$.
Consistent reconstruction is a method for producing an estimate $\widetilde{x} \in \mathbb{R}^d$ of a signal $x\in \mathbb{R}^d$ if one is given a collection of $N$ noisy linear measurements $q_n = \langle x, φ_n \rangle + ε_n$, $1 \leq n \leq N$, that have been corrupted by i.i.d. uniform noise $\{ε_n\}_{n=1}^N$. We prove mean squared error bounds for consistent reconstruction when the measurement vectors $\{φ_n\}_{n=1}^N\subset \mathbb{R}^d$ are drawn independently at random from a suitable distribution on the unit-sphere $\mathbb{S}^{d-1}$. Our main results prove that the mean squared error (MSE) for consistent reconstruction is of the optimal order $\mathbb{E}\|x - \widetilde{x}\|^2 \leq Kδ^2/N^2$ under general conditions on the measurement vectors. We also prove refined MSE bounds when the measurement vectors are i.i.d. uniformly distributed on the unit-sphere $\mathbb{S}^{d-1}$ and, in particular, show that in this case the constant $K$ is dominated by $d^3$, the cube of the ambient dimension. The proofs involve an analysis of random polytopes using coverage processes on the sphere.
Uncertainty principles for generating systems $\{e_n\}_{n=1}^{\infty} \subset \ltwo$ are proven and quantify the interplay between $\ell^r(\N)$ coefficient stability properties and time-frequency localization with respect to $|t|^p$ power weight dispersions. As a sample result, it is proven that if the unit-norm system $\{e_n\}_{n=1}^{\infty}$ is a Schauder basis or frame for $\ltwo$ then the two dispersion sequences $Δ(e_n)$, $Δ(\bar{e_n})$ and the one mean sequence $μ(e_n)$ cannot all be bounded. On the other hand, it is constructively proven that there exists a unit-norm exact system $\{f_n\}_{n=1}^{\infty}$ in $\ltwo$ for which all four of the sequences $Δ(f_n)$, $Δ(\bar{f_n})$, $μ(f_n)$, $μ(\bar{f_n})$ are bounded.
The aim of this paper is to provide complementary quantitative extensions of two results of H.S. Shapiro on the time-frequency concentration of orthonormal sequences in $L^2 (\R)$. More precisely, Shapiro proved that if the elements of an orthonormal sequence and their Fourier transforms are all pointwise bounded by a fixed function in $L^2(\R)$ then the sequence is finite. In a related result, Shapiro also proved that if the elements of an orthonormal sequence and their Fourier transforms have uniformly bounded means and dispersions then the sequence is finite. This paper gives quantitative bounds on the size of the finite orthonormal sequences in Shapiro's uncertainty principles. The bounds are obtained by using prolate spheroïdal wave functions and combinatorial estimates on the number of elements in a spherical code. Extensions for Riesz bases and different measures of time-frequency concentration are also given.