arXiv subjects

Surya T. Tokdar

Publications and source records attributed to Surya T. Tokdar.

At least 19 recordsLinked to original sources

arXiv2026

Adaptive Generalized Elliptical Slice Sampling

A central challenge in gradient-free MCMC is designing algorithms that simultaneously bypass manual tuning, scale efficiently with dimension, and adapt to local target geometry. While adaptive strategies can auto-tune generic frameworks like random walk Metropolis, they offer slow, linear-order scaling of mixing times with dimension. Elliptical slice sampling (ESS) offers a promising alternative: it is tuning-free, adjusts to local geometry, and can achieve nearly dimension-free scaling under favorable conditions. However, its efficiency degrades rapidly if there is a mismatch between the target distribution and the distribution used to generate the ellipse-defining auxiliary variables, precluding its use in high-dimensional settings. We demonstrate that a careful synthesis of ESS and diminishing adaptation directly resolves these bottlenecks. The resulting adaptive generalized elliptical slice sampler (AGESS) self-corrects from a slow-mixing to a fast-mixing regime, while preserving ergodicity across a wide variety of target densities satisfying mild regularity conditions. The algorithm's utility is demonstrated across a broad collection of challenging applications, including generalized regression, deep Gaussian process surrogate modeling, and high-dimensional sparse regression. Together, our theoretical results and the case studies give evidence of the efficiency and robustness of AGESS across target distributions that are non-elliptical, non-differentiable, multi-modal, or high-dimensional.

stat.CO↗

arXiv2025

Modeling Neural Switching via Drift-Diffusion Models

Neural encoding is a field in neuroscience that focuses on characterizing how information from stimuli is encoded in the spiking activity of neurons. When more than one stimulus is present, a theory known as multiplexing posits that neurons temporally switch between encoding various stimuli, creating a fluctuating firing pattern. Here, we propose a new statistical framework to analyze rate fluctuations and discern whether neurons employ multiplexing as a means of encoding multiple stimuli. We adopt a mechanistic approach to modeling multiplexing by constructing a non-Markovian endogenous state-space model. Specifically, we posit that multiplexing arises from competition between the stimuli, which are modeled as latent drift-diffusion processes. We propose a new MCMC algorithm for conducting posterior inference on similar types of state-space models, where typical state-space MCMC methods fail due to strong dependence between the parameters. In addition, we develop alternative models that represent a wide class of alternative encoding theories and perform model comparison using WAIC to determine whether the data suggest the occurrence multiplexing over alternative theories of neural encoding. Using the proposed framework, we provide evidence of multiplexing within the inferior colliculus and novel insight into the switching dynamics.

stat.ME↗

arXiv2025

A Bayesian decision-theoretic approach to sparse estimation

We extend the work of Hahn and Carvalho (2015) and develop a doubly-regularized sparse regression estimator by synthesizing Bayesian regularization with penalized least squares within a decision-theoretic framework. In contrast to existing Bayesian decision-theoretic formulation chiefly reliant upon the symmetric 0-1 loss, the new method -- which we call Bayesian Decoupling -- employs a family of penalized loss functions indexed by a sparsity-tuning parameter. We propose a class of reweighted l1 penalties, with two specific instances that achieve simultaneous bias reduction and convexity. The design of the penalties incorporates considerations of signal sizes, as enabled by the Bayesian paradigm. The tuning parameter is selected using a posterior benchmarking criterion, which quantifies the drop in predictive power relative to the posterior mean which is the optimal Bayes estimator under the squared error loss. Additionally, in contrast to the widely used median probability model technique which selects variables by thresholding posterior inclusion probabilities at the fixed threshold of 1/2, Bayesian Decoupling enables the use of a data-driven threshold which automatically adapts to estimated signal sizes and offers far better performance in high-dimensional settings with highly correlated predictors. Our numerical results in such settings show that certain combinations of priors and loss functions significantly improve the solution path compared to existing methods, prioritizing true signals early along the path before false signals are selected. Consequently, Bayesian Decoupling produces estimates with better prediction and selection performance. Finally, a real data application illustrates the practical advantages of our approaches which select sparser models with larger coefficient estimates.

stat.ME↗

arXiv2024

High-dimensional Bayesian Fourier Analysis For Detecting Circadian Gene Expressions

In genomic applications, there is often interest in identifying genes whose time-course expression trajectories exhibit periodic oscillations with a period of approximately 24 hours. Such genes are usually referred to as circadian, and their identification is a crucial step toward discovering physiological processes that are clock-controlled. It is natural to expect that the expression of gene i at time j might depend to some degree on the expression of the other genes measured at the same time. However, widely-used rhythmicity detection techniques do not accommodate for the potential dependence across genes. We develop a Bayesian approach for periodicity identification that explicitly takes into account the complex dependence structure across time-course trajectories in gene expressions. We employ a latent factor representation to accommodate dependence, while representing the true trajectories in the Fourier domain allows for inference on period, phase, and amplitude of the signal. Identification of circadian genes is allowed through a carefully chosen variable selection prior on the Fourier basis coefficients. The methodology is applied to a novel mouse liver circadian dataset. Although motivated by time-course gene expression array data, the proposed approach is applicable to the analysis of dependent functional data at broad.

stat.ME↗

arXiv2020

Variable Selection Consistency of Gaussian Process Regression

Bayesian nonparametric regression under a rescaled Gaussian process prior offers smoothness-adaptive function estimation with near minimax-optimal error rates. Hierarchical extensions of this approach, equipped with stochastic variable selection, are known to also adapt to the unknown intrinsic dimension of a sparse true regression function. But it remains unclear if such extensions offer variable selection consistency, i.e., if the true subset of important variables could be consistently learned from the data. It is shown here that variable consistency may indeed be achieved with such models at least when the true regression function has finite smoothness to induce a polynomially larger penalty on inclusion of false positive predictors. Our result covers the high dimensional asymptotic setting where the predictor dimension is allowed to grow with the sample size. The proof utilizes Schwartz theory to establish that the posterior probability of wrong selection vanishes asymptotically. A necessary and challenging technical development involves providing sharp upper and lower bounds to small ball probabilities at all rescaling levels of the Gaussian process prior, a result that could be of independent interest.

math.ST↗

arXiv2020

Sensitivity and specificity of a Bayesian single trial analysis for time varying neural signals

We recently reported the existence of fluctuations in neural signals that may permit neurons to code multiple simultaneous stimuli sequentially across time. This required deploying a novel statistical approach to permit investigation of neural activity at the scale of individual trials. Here we present tests using synthetic data to assess the sensitivity and specificity of this analysis. We fabricated datasets to match each of several potential response patterns derived from single-stimulus response distributions. In particular, we simulated dual stimulus trial spike counts that reflected fluctuating mixtures of the single stimulus spike counts, stable intermediate averages, single stimulus winner-take-all, or response distributions that were outside the range defined by the single stimulus responses (such as summation or suppression). We then assessed how well the analysis recovered the correct response pattern as a function of the number of simulated trials and the difference between the simulated responses to each "stimulus" alone. We found excellent recovery of the mixture, intermediate, and outside categories (>97% percent correct), and good recovery of the single/winner-take-all category (>90% correct) when the number of trials was >20 and the single-stimulus response rates were 50Hz and 20Hz respectively. Both larger numbers of trials and greater separation between the single stimulus firing rates improved categorization accuracy. These results provide a benchmark, and guidelines for data collection, for use of this method to investigate coding of multiple items at the individual-trial time scale.

q-bio.NC↗

arXiv2019

Bayesian test of normality versus a Dirichlet process mixture alternative

We propose a Bayesian test of normality for univariate or multivariate data against alternative nonparametric models characterized by Dirichlet process mixture distributions. The alternative models are based on the principles of embedding and predictive matching. They can be interpreted to offer random granulation of a normal distribution into a mixture of normals with mixture components occupying a smaller volume the farther they are from the distribution center. A scalar parametrization based on latent clustering is used to cover an entire spectrum of separation between the normal distributions and the alternative models. An efficient sequential importance sampler is developed to calculate Bayes factors. Simulations indicate the proposed test can detect non-normality without favoring the nonparametric alternative when normality holds.

math.ST↗

arXiv2019

Joint Quantile Regression for Spatial Data

Linear quantile regression is a powerful tool to investigate how predictors may affect a response heterogeneously across different quantile levels. Unfortunately, existing approaches find it extremely difficult to adjust for any dependency between observation units, largely because such methods are not based upon a fully generative model of the data. For analyzing spatially indexed data, we address this difficulty by generalizing the joint quantile regression model of Yang and Tokdar (2017) and characterizing spatial dependence via a Gaussian or $t$ copula process on the underlying quantile levels of the observation units. A Bayesian semiparametric approach is introduced to perform inference of model parameters and carry out spatial quantile smoothing. An effective model comparison criteria is provided, particularly for selecting between different model specifications of tail heaviness and tail dependence. Extensive simulation studies and an application to particulate matter concentration in northeast US are presented to illustrate substantial gains in inference quality, accuracy and uncertainty quantification over existing alternatives.

stat.ME↗

arXiv2016

Paired-move multiple-try stochastic search for Bayesian variable selection

Variable selection is a key issue when analyzing high-dimensional data. The explosion of data with large sample sizes and dimensionality brings new challenges to this problem in both inference accuracy and computational complexity. To alleviate these problems, we propose a new scalable Markov chain Monte Carlo (MCMC) sampling algorithm for "large $p$ small $n$" scenarios by generalizing multiple-try Metropolis to discrete model spaces and further incorporating neighborhood-based stochastic search. The proof of reversibility of the proposed MCMC algorithm is provided. Extensive simulation studies are performed to examine the efficiency of the new algorithm compared with existing methods. A real data example is provided to illustrate the prediction performances of the new algorithm.

stat.ME↗

arXiv2015

Bayesian Analysis of Dynamic Linear Topic Models

In dynamic topic modeling, the proportional contribution of a topic to a document depends on the temporal dynamics of that topic's overall prevalence in the corpus. We extend the Dynamic Topic Model of Blei and Lafferty (2006) by explicitly modeling document level topic proportions with covariates and dynamic structure that includes polynomial trends and periodicity. A Markov Chain Monte Carlo (MCMC) algorithm that utilizes Polya-Gamma data augmentation is developed for posterior inference. Conditional independencies in the model and sampling are made explicit, and our MCMC algorithm is parallelized where possible to allow for inference in large corpora. To address computational bottlenecks associated with Polya-Gamma sampling, we appeal to the Central Limit Theorem to develop a Gaussian approximation to the Polya-Gamma random variable. This approximation is fast and reliable for parameter values relevant in the text mining domain. Our model and inference algorithm are validated with multiple simulation examples, and we consider the application of modeling trends in PubMed abstracts. We demonstrate that sharing information across documents is critical for accurately estimating document-specific topic proportions. We also show that explicitly modeling polynomial and periodic behavior improves our ability to predict topic prevalence at future time points.

stat.ML↗

arXiv2015

Minimax-optimal nonparametric regression in high dimensions

Minimax $L_2$ risks for high-dimensional nonparametric regression are derived under two sparsity assumptions: (1) the true regression surface is a sparse function that depends only on $d=O(\log n)$ important predictors among a list of $p$ predictors, with $\log p=o(n)$; (2) the true regression surface depends on $O(n)$ predictors but is an additive function where each additive component is sparse but may contain two or more interacting predictors and may have a smoothness level different from other components. For either modeling assumption, a practicable extension of the widely used Bayesian Gaussian process regression method is shown to adaptively attain the optimal minimax rate (up to $\log n$ terms) asymptotically as both $n,p\to\infty$ with $\log p=o(n)$.

math.ST↗

arXiv2015

Computer emulation with non-stationary Gaussian processes

Gaussian process (GP) models are widely used to emulate propagation uncertainty in computer experiments. GP emulation sits comfortably within an analytically tractable Bayesian framework. Apart from propagating uncertainty of the input variables, a GP emulator trained on finitely many runs of the experiment also offers error bars for response surface estimates at unseen input values. This helps select future input values where the experiment should be run to minimize the uncertainty in the response surface estimation. However, traditional GP emulators use stationary covariance functions, which perform poorly and lead to sub-optimal selection of future input points when the response surface has sharp local features, such as a jump discontinuity or an isolated tall peak. We propose an easily implemented non-stationary GP emulator, based on two stationary GPs, one nested into the other, and demonstrate its superior ability in handling local features and selecting future input points from the boundaries of such features.

stat.ME↗

arXiv2014

Additive Gaussian Process Regression

Additive-interactive regression has recently been shown to offer attractive minimax error rates over traditional nonparametric multivariate regression in a wide variety of settings, including cases where the predictor count is much larger than the sample size and many of the predictors have important effects on the response, potentially through complex interactions. We present a Bayesian implementation of additive-interactive regression using an additive Gaussian process (AGP) prior and develop an efficient Markov chain sampler that extends stochastic search variable selection in this setting. Careful prior and hyper-parameter specification are developed in light of performance and computational considerations, and key innovations address difficulties in exploring a joint posterior distribution over multiple subsets of high dimensional predictor inclusion vectors. The method offers state-of-the-art support and interaction recovery while improving dramatically over competitors in terms of prediction accuracy on a diverse set of simulated and real data. Results from real data studies provide strong evidence that the additive-interactive framework is an attractive modeling platform for high-dimensional nonparametric regression.

stat.ME↗

arXiv2013

Adaptive Bayesian multivariate density estimation with Dirichlet mixtures

We show that rate-adaptive multivariate density estimation can be performed using Bayesian methods based on Dirichlet mixtures of normal kernels with a prior distribution on the kernel's covariance matrix parameter. We derive sufficient conditions on the prior specification that guarantee convergence to a true density at a rate that is optimal minimax for the smoothness class to which the true density belongs. No prior knowledge of smoothness is assumed. The sufficient conditions are shown to hold for the Dirichlet location mixture of normals prior with a Gaussian base measure and an inverse-Wishart prior on the covariance matrix parameter. Locally Hölder smoothness classes and their anisotropic extensions are considered. Our study involves several technical novelties, including sharp approximation of finitely differentiable multivariate densities by normal mixtures and a new sieve on the space of such densities.

math.ST↗

arXiv2011

Dimension adaptability of Gaussian process models with variable selection and projection

It is now known that an extended Gaussian process model equipped with rescaling can adapt to different smoothness levels of a function valued parameter in many nonparametric Bayesian analyses, offering a posterior convergence rate that is optimal (up to logarithmic factors) for the smoothness class the true function belongs to. This optimal rate also depends on the dimension of the function's domain and one could potentially obtain a faster rate of convergence by casting the analysis in a lower dimensional subspace that does not amount to any loss of information about the true function. In general such a subspace is not known a priori but can be explored by equipping the model with variable selection or linear projection. We demonstrate that for nonparametric regression, classification, density estimation and density regression, a rescaled Gaussian process model equipped with variable selection or linear projection offers a posterior convergence rate that is optimal (up to logarithmic factors) for the lowest dimension in which the analysis could be cast without any loss of information about the true function. Theoretical exploration of such dimension reduction features appears novel for Bayesian nonparametric models with or without Gaussian processes.

math.ST↗

arXiv2011

Adaptive Convergence Rates of a Dirichlet Process Mixture of Multivariate Normals

It is shown that a simple Dirichlet process mixture of multivariate normals offers Bayesian density estimation with adaptive posterior convergence rates. Toward this, a novel sieve for non-parametric mixture densities is explored, and its rate adaptability to various smoothness classes of densities in arbitrary dimension is demonstrated. This sieve construction is expected to offer a substantial technical advancement in studying Bayesian non-parametric mixture models based on stick-breaking priors.

math.ST↗

arXiv2011

A nonparametric empirical Bayes framework for large-scale multiple testing

We propose a flexible and identifiable version of the two-groups model, motivated by hierarchical Bayes considerations, that features an empirical null and a semiparametric mixture model for the non-null cases. We use a computationally efficient predictive recursion marginal likelihood procedure to estimate the model parameters, even the nonparametric mixing distribution. This leads to a nonparametric empirical Bayes testing procedure, which we call PRtest, based on thresholding the estimated local false discovery rates. Simulations and real-data examples demonstrate that, compared to existing approaches, PRtest's careful handling of the non-null density can give a much better fit in the tails of the mixture distribution which, in turn, can lead to more realistic conclusions.

stat.ME↗

arXiv2011

Semiparametric inference in mixture models with predictive recursion marginal likelihood

Predictive recursion is an accurate and computationally efficient algorithm for nonparametric estimation of mixing densities in mixture models. In semiparametric mixture models, however, the algorithm fails to account for any uncertainty in the additional unknown structural parameter. As an alternative to existing profile likelihood methods, we treat predictive recursion as a filter approximation to fitting a fully Bayes model, whereby an approximate marginal likelihood of the structural parameter emerges and can be used for inference. We call this the predictive recursion marginal likelihood. Convergence properties of predictive recursion under model mis-specification also lead to an attractive construction of this new procedure. We show pointwise convergence of a normalized version of this marginal likelihood function. Simulations compare the performance of this new marginal likelihood approach that of existing profile likelihood methods as well as Dirichlet process mixtures in density estimation. Mixed-effects models and an empirical Bayes multiple testing application in time series analysis are also considered.

stat.ME↗