arXiv subjects

Nariankadu D. Shyamalkumar

Publications and source records attributed to Nariankadu D. Shyamalkumar.

3 recordsLinked to original sources

arXiv2020

An Algorithm for Distributed Bayesian Inference in Generalized Linear Models

Monte Carlo algorithms, such as Markov chain Monte Carlo (MCMC) and Hamiltonian Monte Carlo (HMC), are routinely used for Bayesian inference in generalized linear models; however, these algorithms are prohibitively slow in massive data settings because they require multiple passes through the full data in every iteration. Addressing this problem, we develop a scalable extension of these algorithms using the divide-and-conquer (D&C) technique that divides the data into a sufficiently large number of subsets, draws parameters in parallel on the subsets using a \textit{powered} likelihood, and produces Monte Carlo draws of the parameter by combining parameter draws obtained from each subset. These combined parameter draws play the role of draws from the original sampling algorithm. Our main contributions are two-fold. First, we demonstrate through diverse simulated and real data analyses that our distributed algorithm is comparable to the current state-of-the-art D&C algorithm in terms of statistical accuracy and computational efficiency. Second, providing theoretical support for our empirical observations, we identify regularity assumptions under which the proposed algorithm leads to asymptotically optimal inference. We illustrate our methodology through normal linear and logistic regressions, where parts of our D&C algorithm are analytically tractable.

stat.CO↗

arXiv2019

On Tail Dependence Matrices -- The Realization Problem for Parametric Families

Among bivariate tail dependence measures, the tail dependence coefficient has emerged as the popular choice. Akin to the correlation matrix, a multivariate dependence measure is constructed using these bivariate measures, and this is referred to in the literature as the tail dependence matrix (TDM). While the problem of determining whether a given ${d\times d}$ matrix is a correlation matrix is of the order $O(d^3)$ in complexity, determining if a matrix is a TDM (the realization problem) is believed to be non-polynomial in complexity. Using a linear programming (LP) formulation, we show that the combinatorial structure of the constraints is related to the intractable max-cut problem in a weighted graph. This connection provides an avenue for constructing parametric classes admitting a polynomial in $d$ algorithm for determining membership in its constraint polytope. The complexity of the general realization problem is justifiably of much theoretical interest. Since in practice one resorts to lower dimensional parametrization of the TDMs, we posit that it is rather the complexity of the realization problem restricted to parametric classes of TDMs, and algorithms for it, that are more practically relevant. In this paper, we show how the inherent symmetry and sparsity in a parametrization can be exploited to achieve a significant reduction in the LP formulation, which can lead to polynomial complexity of such realization problems - some parametrizations even resulting in the constraint polytope being independent of $d$. We also explore the use of a probabilistic viewpoint on TDMs to derive the constraint polytopes.

math.ST↗

arXiv2007

Reading policies for joins: An asymptotic analysis

Suppose that $m_n$ observations are made from the distribution $\mathbf {R}$ and $n-m_n$ from the distribution $\mathbf {S}$. Associate with each pair, $x$ from $\mathbf {R}$ and $y$ from $\mathbf {S}$, a nonnegative score $ϕ(x,y)$. An optimal reading policy is one that yields a sequence $m_n$ that maximizes $\mathbb{E}(M(n))$, the expected sum of the $(n-m_n)m_n$ observed scores, uniformly in $n$. The alternating policy, which switches between the two sources, is the optimal nonadaptive policy. In contrast, the greedy policy, which chooses its source to maximize the expected gain on the next step, is shown to be the optimal policy. Asymptotics are provided for the case where the $\mathbf {R}$ and $\mathbf {S}$ distributions are discrete and $ϕ(x,y)=1 or 0$ according as $x=y$ or not (i.e., the observations match). Specifically, an invariance result is proved which guarantees that for a wide class of policies, including the alternating and the greedy, the variable M(n) obeys the same CLT and LIL. A more delicate analysis of the sequence $\mathbb{E}(M(n))$ and the sample paths of M(n), for both alternating and greedy, reveals the slender sense in which the latter policy is asymptotically superior to the former, as well as a sense of equivalence of the two and robustness of the former.

math.PR↗