# Publications

We present the optimal design of a spectral method widely used to initialize nonconvex optimization algorithms for solving phase retrieval and other signal recovery problems. Our work leverages recent results that provide an exact characterization of the performance of the spectral method in the high-dimensional limit. This characterization allows us to map the task of optimal design to a constrained optimization problem in a weighted $L^2$ function space. The latter has a closed-form solution. Interestingly, under a mild technical condition, our results show that there exists a fixed design that is uniformly optimal over all sampling ratios. Numerical simulations demonstrate the performance improvement brought by the proposed optimal design over existing constructions in the literature. In a recent work, Mondelli and Montanari have shown the existence of a weak reconstruction threshold below which the spectral method cannot provide useful estimates. Our results serve to complement that work by deriving the fundamental limit of the spectral method beyond the aforementioned threshold.

Substantial progress has been made recently on developing provably accurate and efficient algorithms for low-rank matrix factorization via nonconvex optimization. While conventional wisdom often takes a dim view of nonconvex optimization algorithms due to their susceptibility to spurious local minima, simple iterative methods such as gradient descent have been remarkably successful in practice. The theoretical footings, however, had been largely lacking until recently.

In this tutorial-style overview, we highlight the important role of statistical models in enabling efficient nonconvex optimization with performance guarantees. We review two contrasting approaches: (1) two-stage algorithms, which consist of a tailored initialization step followed by successive refinement; and (2) global landscape analysis and initialization-free algorithms. Several canonical matrix factorization problems are discussed, including but not limited to matrix sensing, phase retrieval, matrix completion, blind deconvolution, robust principal component analysis, phase synchronization, and joint alignment. Special care is taken to illustrate the key technical insights underlying their analyses. This article serves as a testament that the integrated thinking of optimization and statistics leads to fruitful research findings.

We study the problem of detecting a random walk on a graph from a sequence of noisy measurements at every node. There are two hypotheses: either every observation is just meaningless zero-mean Gaussian noise, or at each time step exactly one node has an elevated mean, with its location following a random walk on the graph over time. We want to exploit knowledge of the graph structure and random walk parameters (specified by a Markov chain transition matrix) to detect a possibly very weak signal. The optimal detector is easily derived, and we focus on the harder problem of characterizing its performance through the (type-II) error exponent: the decay rate of the miss probability under a false alarm constraint.

The expression for the error exponent resembles the free energy of a spin glass in statistical physics, and we borrow techniques from that field to develop a lower bound. Our fully rigorous analysis uses large deviations theory to show that the lower bound exhibits a phase transition: strong performance is only guaranteed when the signal-to-noise ratio exceeds twice the entropy rate of the random walk.

Monte Carlo simulations show that the lower bound fully captures the behavior of the true exponent.

We study a spectral initialization method that serves as a key ingredient in recent work on using efficient iterative algorithms for estimating signals in nonconvex settings. Unlike previous analysis in the literature, which is restricted to the phase retrieval setting and which provides only performance bounds, we consider arbitrary generalized linear sensing models and present a precise asymptotic characterization of the performance of the spectral method in the high-dimensional regime. Our analysis reveals a phase transition phenomenon that depends on the sampling ratio. When the ratio is below a minimum threshold, the estimates given by the spectral method are no better than a random guess drawn uniformly from the hypersphere; above a maximum threshold, however, the estimates become increasingly aligned with the target signal. The computational complexity of the spectral method is also markedly different in the two phases. Worked examples and numerical results are provided to illustrate and verify the analytical predictions. In particular, simulations show that our asymptotic formulas provide accurate predictions even at moderate signal dimensions.

Finding the sparse representation of a signal in an overcomplete dictionary has attracted a lot of attention over the past years. This paper studies ProSparse, a new polynomial complexity algorithm that solves the sparse representation problem when the underlying dictionary is the union of a Vandermonde matrix and a banded matrix. Unlike our previous work which establishes deterministic (worst-case) sparsity bounds for ProSparse to succeed, this paper presents a probabilistic average-case analysis of the algorithm. Based on a generating-function approach, closed-form expressions for the exact success probabilities of ProSparse are given. The success probabilities are also analyzed in the high-dimensional regime. This asymptotic analysis characterizes a sharp phase transition phenomenon regarding the performance of the algorithm.

Sparse coding refers to the pursuit of the sparsest representation of a signal in a typically overcomplete dictionary. From a Bayesian perspective, sparse coding provides a Maximum a Posteriori (MAP) estimate of the unknown vector under a sparse prior. Various nonlinear algorithms are available to approximate the solution of such problems.

In this work, we suggest enhancing the performance of sparse coding algorithms by a deliberate and controlled contamination of the input with random noise, a phenomenon known as stochastic resonance. This not only allows for increased performance, but also provides a computationally efficient approximation to the Minimum Mean Square Error (MMSE) estimator, which is ordinarily intractable to compute. We demonstrate our findings empirically and provide a theoretical analysis of our method under several different cases.

We propose an image representation scheme combining the local and nonlocal characterization of patches in an image. Our representation scheme can be shown to be equivalent to a tight frame constructed from convolving local bases (e.g., wavelet frames, discrete cosine transforms, etc.) with nonlocal bases (e.g., spectral basis induced by nonlinear dimension reduction on patches), and we call the resulting frame elements convolution framelets. Insight gained from analyzing the proposed representation leads to a novel interpretation of a recent high-performance patch-based image pro- cessing algorithm using the point integral method (PIM) and the low dimensional manifold model (LDMM) [S. Osher, Z. Shi, and W. Zhu, Low Dimensional Manifold Model for Image Processing, Tech. Rep., CAM report 16-04, UCLA, Los Angeles, CA, 2016]. In particular, we show that LDMM is a weighted l2-regularization on the coefficients obtained by decomposing images into linear combinations of convolution framelets; based on this understanding, we extend the original LDMM to a reweighted version that yields further improved results. In addition, we establish the energy concentration property of convolution framelet coefficients for the setting where the local basis is constructed from a given nonlocal basis via a linear reconstruction framework; a generalization of this framework to unions of local embeddings can provide a natural setting for interpreting BM3D, one of the state-of-the-art image denoising algorithms.

Many patch-based image denoising algorithms can be formulated as applying a smoothing filter to the noisy image. Expressed as matrices, the smoothing filters must be row normalized so that each row sums to unity. Surprisingly, if we apply a column normalization before the row normalization, the performance of the smoothing filter can often be significantly improved. Prior works showed that such performance gain is related to the Sinkhorn-Knopp balancing algorithm, an iterative procedure that symmetrizes a row-stochastic matrix to a doubly-stochastic matrix. However, a complete understanding of the performance gain phenomenon is still lacking.

In this paper, we study the performance gain phenomenon from a statistical learning perspective. We show that Sinkhorn-Knopp is equivalent to an Expectation-Maximization (EM) algorithm of learning a Product of Gaussians (PoG) prior of the image patches. By establishing the correspondence between the steps of Sinkhorn-Knopp and the EM algorithm, we provide a geometrical interpretation of the symmetrization process. The new PoG model also allows us to develop a new denoising algorithm called Product of Gaussian Non-Local-Means (PoG-NLM). PoG-NLM is an extension of the Sinkhorn-Knopp and is a generalization of the classical non-local means. Despite its simple formulation, PoG-NLM outperforms many existing smoothing filters and has a similar performance compared to BM3D.

Estimating low-rank positive-semidefinite (PSD) matrices from symmetric rank-one measurements is of great importance in many applications, such as high-dimensional data processing, quantum state tomography, and phase retrieval. When the rank is known a priori, this problem can be regarded as solving a system of quadratic equations of a low-dimensional subspace. The authors develop a fast iterative algorithm based on an adaptation of the Kaczmarz method, which is traditionally used for solving overdetermined linear systems. In particular, the authors characterize the dynamics of the algorithm when the measurement vectors are composed of standard Gaussian entries in the online setting. Numerical simulations demonstrate the compelling performance of the proposed algorithm.

Standard convolution as a model of radiometric degradation is in majority of cases inaccurate as the blur varies in space and we are thus required to work with a computationally demanding space-variant model. Space-variant degradation can be approximately decomposed to a set of standard convolutions. We explain in detail the properties of the space-variant degrada- tion operator and show two possible decomposition models and two approximation approaches. Our target application is space- variant image deconvolution, on which we illustrate theoretical differences between these models. We propose a computationally efficient restoration algorithm that belongs to a category of alternating direction methods of multipliers, which consists of four update steps with closed-form solutions. Depending on the used decomposition, two variations of the algorithm exist with distinct properties. We test the effectiveness of the decomposition models under different levels of approximation on synthetic and real examples, and conclude the letter by drawing several practical observations.

Perceptual decision-making is fundamental to a broad range of fields including neurophysiology, economics, medicine, advertising, law, etc. While recent findings have yielded major advances in our understanding of perceptual decision-making, decision-making as a function of time and frequency (i.e., decision-making dynamics) is not well understood. To limit the review length, we focus most of this review on human findings. Animal findings, which are extensively reviewed elsewhere, are included when beneficial or necessary. We attempt to put these various findings and datasets - which can appear to be unrelated in the absence of a formal dynamic analysis - into context using published models. Specifically, by adding appropriate dynamic mechanisms (e.g., high-pass filters) to existing models, it appears that a number of otherwise seemingly disparate findings from the literature might be explained. One hypothesis that arises through this dynamic analysis is that decision-making includes phasic (high-pass) neural mechanisms, an evidence accumulator and/or some sort of mid-trial decision-making mechanism (e.g., peak detector and/or decision boundary).

Motivated by recent progresses in signal processing on graphs, we have developed a matched signal detection (MSD) theory for signals with intrinsic structures described by weighted graphs. First, we regard graph Laplacian eigenvalues as frequencies of graph-signals and assume that the signal is in a subspace spanned by the first few graph Laplacian eigenvectors associated with lower eigenvalues. The conventional matched subspace detector can be applied to this case. Furthermore, we study signals that may not merely live in a subspace. Namely, we consider signals with bounded variation on graphs and more general signals that are randomly drawn from a prior distribution. For bounded variation signals, the test is a weighted energy detector. For the random signals, the test statistic is the difference of signal variations on associated graphs, if a degenerate Gaussian distribution specified by the graph Laplacian is adopted. We evaluate the effectiveness of the MSD on graphs both on simulated and real data sets. Specifically, we apply MSD to the brain imaging data classification problem of Alzheimer’s disease (AD) based on two independent data sets: 1) positron emission tomography data with Pittsburgh compound-B tracer of 30 AD and 40 normal control (NC) subjects, 2) resting-state functional magnetic resonance imaging (R-fMRI) data of 30 early mild cognitive impairment and 20 NC subjects. Our results demonstrate that the MSD approach is able to outperform the traditional methods and help detect AD at an early stage, probably due to the success of exploiting the manifold structure of the data.

We propose a fully distributed Gauss-Newton algorithm for state estimation of electric power systems. At each Gauss-Newton iteration, matrix-splitting techniques are utilized to carry out the matrix inversion needed for calculating the Gauss-Newton step in a distributed fashion. In order to reduce the communication burden as well as increase robustness of state estimation, the proposed distributed scheme relies only on local information and a limited amount of information from neighboring areas. The matrix-splitting scheme is designed to calculate the Gauss-Newton step with exponential convergence speed. The effectiveness of the method is demonstrated in various numerical experiments.

We propose a sampling scheme that can perfectly reconstruct a collection of

spikes on the sphere from samples of their lowpass-filtered observations.

Central to our algorithm is a generalization of the annihilating filter

method, a tool widely used in array signal processing and finite-rate-of-innovation

(FRI) sampling. The proposed algorithm can reconstruct $K$ spikes

from $(K+\sqrt{K})^2$ spatial samples---a sampling requirement that

improves over known sparse sampling schemes on the sphere by a factor of up

to four.

We showcase the versatility of the proposed algorithm by applying it to

three different problems: 1) sampling diffusion processes induced by

localized sources on the sphere, 2) shot-noise removal, and 3) sound source

localization (SSL) by a spherical microphone array. In particular, we show

how SSL can be reformulated as a spherical sparse sampling problem.

Understanding network features of brain pathology is essential to reveal underpinnings of neurodegenerative diseases. In this paper, we introduce a novel graph regression model (GRM) for learning structural brain connectivity of Alzheimer’s disease (AD) measured by amyloid-β deposits. The proposed GRM regards 11 C-labeled Pittsburgh Compound-B (PiB) positron emission tomography (PET) imaging data as smooth signals defined on an unknown graph. This graph is then estimated through an optimization framework, which fits the graph to the data with an adjustable level of uniformity of the connection weights. Under the assumed data model, results based on simulated data illustrate that our approach can accurately reconstruct the underlying network, often with better reconstruction than those obtained by both sample correlation and ℓ1 -regularized partial correlation estimation. Evaluations performed upon PiB-PET imaging data of 30 AD and 40 elderly normal control (NC) subjects demonstrate that the connectivity patterns revealed by the GRM are easy to interpret and consistent with known pathology. Moreover, the hubs of the reconstructed networks match the cortical hubs given by functional MRI. The discriminative network features including both global connectivity measurements and degree statistics of specific nodes discovered from the AD and NC amyloid-beta networks provide new potential biomarkers for preclinical and clinical AD.

We propose a randomized version of the non-local means (NLM) algorithm for large-scale image filtering. The new algorithm, called Monte Carlo non-local means (MCNLM), speeds up the classical NLM by computing a small subset of image patch distances, which are randomly selected according to a designed sampling pattern. We make two contributions. First, we analyze the performance of the MCNLM algorithm and show that, for large images or large external image databases, the random outcomes of MCNLM are tightly concentrated around the deterministic full NLM result. In particular, our error probability bounds show that, at any given sampling ratio, the probability for MCNLM to have a large deviation from the original NLM solution decays exponentially as the size of the image or database grows. Second, we derive explicit formulas for optimal sampling patterns that minimize the error probability bound by exploiting partial knowledge of the pairwise similarity weights. Numerical experiments show that MCNLM is competitive with other state-of-the-art fast NLM algorithms for single-image denoising. When applied to denoising images using an external database containing ten billion patches, MCNLM returns a randomized solution that is within 0.2 dB of the full NLM solution while reducing the runtime by three orders of magnitude.