arXiv subjects

Arun K Kuchibhotla

Publications and source records attributed to Arun K Kuchibhotla.

5 recordsLinked to original sources

arXiv2026

Statistical Inference for Online Algorithms

The construction of confidence intervals and hypothesis tests for functionals is a cornerstone of statistical inference. Traditionally, the most efficient procedures - such as the Wald interval or the Likelihood Ratio Test - require both a point estimator and a consistent estimate of its asymptotic variance. However, when estimators are derived from online or sequential algorithms, computational constraints often preclude multiple passes over the data, complicating variance estimation. In this article, we propose a computationally efficient, rate-optimal wrapper method (HulC) that wraps around any online algorithm to produce asymptotically valid confidence regions bypassing the need for explicit asymptotic variance estimation. The method is provably valid for any online algorithm that yields an asymptotically normal estimator. We evaluate the practical performance of the proposed method primarily using Stochastic Gradient Descent (SGD) with Polyak-Ruppert averaging. Furthermore, we provide extensive numerical simulations comparing the performance of our approach (HulC) when used with other online algorithms, including implicit-SGD and ROOT-SGD.

stat.ML↗

arXiv2025

Maximal Inequalities for Independent Random Vectors

Maximal inequalities refer to bounds on expected values of the supremum of averages of random variables over a collection. They play a crucial role in the study of non-parametric and high-dimensional estimators, and especially in the study of empirical risk minimizers. Although the expected supremum over an infinite collection appears more often in these applications, the expected supremum over a finite collection is a basic building block. This follows from the generic chaining argument. For the case of finite maximum, most existing bounds stem from the Bonferroni inequality (or the union bound). The optimality of such bounds is not obvious, especially in the context of heavy-tailed random vectors. In this article, we consider the problem of finding sharp upper and lower bounds for the expected $L_{\infty}$ norm of the mean of finite-dimensional random vectors under marginal variance bounds and an integrable envelope condition.

math.PR↗

arXiv2023

Rectangular Hull Confidence Regions for Multivariate Parameters

We introduce three notions of multivariate median bias, namely, rectilinear, Tukey, and orthant median bias. Each of these median biases is zero under a suitable notion of multivariate symmetry. We study the coverage probabilities of rectangular hull of $B$ independent multivariate estimators, with special attention to the number of estimators $B$ needed to ensure a miscoverage of at most $α$. It is proved that for estimators with zero orthant median bias, we need $B\geq c\log_2(d/α)$ for some constant $c > 0$. Finally, we show that there exists an asymptotically valid (non-trivial) confidence region for a multivariate parameter $θ_0$ if and only if there exists a (non-trivial) estimator with an asymptotic orthant median bias of zero.

math.ST↗

arXiv2023

Central Limit Theorems and Approximation Theory: Part I

Central limit theorems (CLTs) have a long history in probability and statistics. They play a fundamental role in constructing valid statistical inference procedures. Over the last century, various techniques have been developed in probability and statistics to prove CLTs under a variety of assumptions on random variables. Quantitative versions of CLTs (e.g., Berry--Esseen bounds) have also been parallelly developed. In this article, we propose to use approximation theory from functional analysis to derive explicit bounds on the difference between expectations of functions.

math.ST↗

arXiv2020

First order expansion of convex regularized estimators

We consider first order expansions of convex penalized estimators in high-dimensional regression problems with random designs. Our setting includes linear regression and logistic regression as special cases. For a given penalty function $h$ and the corresponding penalized estimator $\hatβ$, we construct a quantity $η$, the first order expansion of $\hatβ$, such that the distance between $\hatβ$ and $η$ is an order of magnitude smaller than the estimation error $\|\hatβ - β^*\|$. In this sense, the first order expansion $η$ can be thought of as a generalization of influence functions from the mathematical statistics literature to regularized estimators in high-dimensions. Such first order expansion implies that the risk of $\hatβ$ is asymptotically the same as the risk of $η$ which leads to a precise characterization of the MSE of $\hatβ$; this characterization takes a particularly simple form for isotropic design. Such first order expansion also leads to inference results based on $\hatβ$. We provide sufficient conditions for the existence of such first order expansion for three regularizers: the Lasso in its constrained form, the lasso in its penalized form, and the Group-Lasso. The results apply to general loss functions under some conditions and those conditions are satisfied for the squared loss in linear regression and for the logistic loss in the logistic model.

math.ST↗