arXiv subjects

Christos Merkatas

Publications and source records attributed to Christos Merkatas.

7 recordsLinked to original sources

arXiv2026

Exact two-stage finite-mixture representations for species sampling processes

Discrete random probability measures are central to Bayesian inference, particularly as priors for mixture modeling and clustering. A broad and unifying class is that of proper species sampling processes (SSPs), encompassing many Bayesian nonparametric priors. We show that any proper SSP admits an exact two-stage finite-mixture representation built from a latent truncation index and a simple reweighting of the atoms. For each realized truncation index, the representation has finitely many atoms, and averaging over the induced law of that index recovers the original SSP setwise. This yields at least two consequences: (i) an exact two-stage finite construction for arbitrary SSPs, without user-chosen truncation levels; and (ii) posterior inference in SSP mixture models via standard finite-mixture machinery, leading to tractable MCMC algorithms without ad hoc truncations. We explore these consequences by deriving explicit total-variation bounds for the approximation error when the truncation level is fixed, and by studying practical performance in mixture modeling, with emphasis on Dirichlet and geometric SSPs.

stat.ME↗

arXiv2022

System identification using Bayesian neural networks with nonparametric noise models

System identification is of special interest in science and engineering. This article is concerned with a system identification problem arising in stochastic dynamic systems, where the aim is to estimate the parameters of a system along with its unknown noise processes. In particular, we propose a Bayesian nonparametric approach for system identification in discrete time nonlinear random dynamical systems assuming only the order of the Markov process is known. The proposed method replaces the assumption of Gaussian distributed error components with a highly flexible family of probability density functions based on Bayesian nonparametric priors. Additionally, the functional form of the system is estimated by leveraging Bayesian neural networks which also leads to flexible uncertainty quantification. Asymptotically on the number of hidden neurons, the proposed model converges to full nonparametric Bayesian regression model. A Gibbs sampler for posterior inference is proposed and its effectiveness is illustrated on simulated and real time series.

stat.ME↗

arXiv2021

On the approximation of basins of attraction using deep neural networks

The basin of attraction is the set of initial points that will eventually converge to some attracting set. Its knowledge is important in understanding the dynamical behavior of a given dynamical system of interest. In this work, we address the problem of reconstructing the basins of attraction of a multistable system, using only labeled data. To this end, we view this problem as a classification task and use a deep neural network as a classifier for predicting the attractor that corresponds to any given initial condition. Additionally, we provide a method for obtaining an approximation of the basin boundary of the underlying system, using the trained classification model. Finally, we provide evidence relating the complexity of the structure of the basins of attraction with the quality of the obtained reconstructions, via the concept of basin entropy. We demonstrate the application of the proposed method on the Lorenz system in a bistable regime.

math.DS↗

arXiv2019

Shades of Dark Uncertainty and Consensus Value for the Newtonian Constant of Gravitation

The Newtonian constant of gravitation, $G$, stands out in the landscape of the most common fundamental constants owing to its surprisingly large relative uncertainty, which is attributable mostly to the dispersion of the values measured for it in different experiments. This study focuses on a set of measurements of $G$ that are mutually inconsistent, in the sense that the dispersion of the measured values is significantly larger than what their reported uncertainties suggest that it should be. Furthermore, there is a loosely defined group of measured values that lie fairly close to a consensus value that may be derived from all the measurement results, and then there are one or more groups with measured values farther away from the consensus value, some higher, others lower. This same general pattern is often observed in many interlaboratory studies and meta-analyses. In the conventional treatments of such data, the mutual inconsistency is addressed by inflating the reported uncertainties, either multiplicatively, or by the addition of random effects, both reflecting the presence of dark uncertainty. The former approach is often used by CODATA and by the Particle Data Group, and the latter is common in medical meta-analysis and in metrology. We propose a new procedure for consensus building that models the results using latent clusters with different shades of dark uncertainty, which assigns a customized amount of dark uncertainty to each measured value, as a mixture of those shades, and does so taking into account both the placement of the measured values relative to the consensus value, and the reported uncertainties. We demonstrate this procedure by deriving a new estimate for $G$, as a consensus value $G = 6.67408 \times 10^{-11} \,\text{m}^{-3} \, \text{kg}^{-1} \, \text{s}^{-2}$, with $u(G) = 0.00024 \times 10^{-11} \,\text{m}^{-3} \, \text{kg}^{-1} \, \text{s}^{-2}$.

physics.data-an↗

arXiv2019

Joint reconstruction and prediction of random dynamical systems under borrowing of strength

We propose a Bayesian nonparametric model based on Markov Chain Monte Carlo (MCMC) methods for the joint reconstruction and prediction of discrete time stochastic dynamical systems, based on $m$-multiple time-series data, perturbed by additive dynamical noise. We introduce the Pairwise Dependent Geometric Stick-Breaking Reconstruction (PD-GSBR) model, which relies on the construction of a $m$-variate nonparametric prior over the space of densities supported over $\mathbb{R}^m$. We are focusing in the case where at least one of the time-series has a sufficiently large sample size representation for an independent and accurate Geometric Stick-Breaking estimation, as defined in Merkatas et al. (2017). Our contention, is that whenever the dynamical error processes perturbing the underlying dynamical systems share common characteristics, underrepresented data sets can benefit in terms of model estimation accuracy. The PD-GSBR estimation and prediction procedure is demonstrated specifically in the case of maps with polynomial nonlinearities of an arbitrary degree. Simulations based on synthetic time-series are presented.

stat.ME↗

arXiv2017

A Bayesian Nonparametric approach to Reconstruction and Prediction of Random Dynamical Systems

We propose a Bayesian nonparametric mixture model for the reconstruction and prediction from observed time series data, of discretized stochastic dynamical systems, based on Markov Chain Monte Carlo methods (MCMC). Our results can be used by researchers in physical modeling interested in a fast and accurate estimation of low dimensional stochastic models when the size of the observed time series is small and the noise process (perhaps) is non-Gaussian. The inference procedure is demonstrated specifically in the case of polynomial maps of arbitrary degree and when a Geometric Stick Breaking mixture process prior over the space of densities, is applied to the additive errors. Our method is parsimonious compared to Bayesian nonparametric techniques based on Dirichlet process mixtures, flexible and general. Simulations based on synthetic time series are presented.

stat.AP↗

arXiv2017

Dependent Mixtures of Geometric Weights Priors

A new approach on the joint estimation of partially exchangeable observations is presented by constructing pairwise dependence between $m$ random density functions, each of which is modeled as a mixture of geometric stick breaking processes. This approach is based on a new random central masses version of the Pairwise Dependent Dirichlet Process prior mixture model (PDDP) first introduced in Hatjispyros et al. (2011). The idea is to create pairwise dependence through random measures that are location-preserving-expectations of Dirichlet random measures. Our contention is that mixture modeling with Pairwise Dependent Geometric Stick Breaking Process (PDGSBP) priors is sufficient for prediction and estimation purposes; moreover the associated Gibbs sampler is much faster and easier to implement than its Dirichlet Process based counterpart. To this respect, we provide a-priori-synchronized comparison studies under sparse $m$-scalable synthetic and real data examples.

stat.ME↗