arXiv subjects

Hock Peng Chan

Publications and source records attributed to Hock Peng Chan.

At least 19 recordsLinked to original sources

arXiv2023

Thresholding the higher criticism test statistics for optimality in a heterogeneous setting

Donoho and Kipnis (2022) showed that the the higher criticism (HC) test statistic has a non-Gaussian phase transition but remarked that it is probably not optimal, in the detection of sparse differences between two large frequency tables when the counts are low. The setting can be considered to be heterogeneous, with cells containing larger total counts more able to detect smaller differences. We provide a general study here of sparse detection arising from such heterogeneous settings, and showed that optimality of the HC test statistic requires thresholding, for example in the case of frequency table comparison, to restrict to p-values of cells with total counts exceeding a threshold. The use of thresholding also leads to optimality of the HC test statistic when it is applied on the sparse Poisson means model of Arias-Castro and Wang (2015). The phase transitions we consider here are non-Gaussian, and involve an interplay between the rate functions of the response and sample size distributions. We also showed, both theoretically and in a numerical study, that applying thresholding to the Bonferroni test statistic results in better sparse mixture detection in heterogeneous settings.

math.ST↗

arXiv2023

Minimizing Change-Point Estimation Error

In this paper we consider change-points in multiple sequences with the objective of minimizing the estimation error of a sequence by making use of information from other sequences. This is in contrast to recent interest on change-points in multiple sequences where the focus is on detection of common change-points. We start with the canonical case of a single sequence with constant change-point intensities. We consider two measures of a change-point algorithm. The first is the probability of estimating the change-point with no error. The second is the expected distance between the true and estimated change-points. We provide a theoretical upper bound for the no error probability, and a lower bound for the expected distance, that must be satisfied by all algorithms. We propose a scan-CUSUM algorithm that achieves the no error upper bound and come close to the distance lower bound. We next consider the case of non-constant intensities and establish sharp conditions under which estimation error can go to zero. We propose an extension of the scan-CUSUM algorithm for a non-constant intensity function, and show that it achieves asymptotically zero error at the boundary of the zero-error regime. We illustrate an application of the scan-CUSUM algorithm on multiple sequences sharing an unknown, non-constant intensity function. We estimate the intensity function from the change-point profile likelihoods of all sequences and apply scan-CUSUM on the estimated intensity function.

math.ST↗

arXiv2022

Likelihood Scores for Sparse Signal and Change-Point Detection

We consider here the identification of change-points on large-scale data streams. The objective is to find the most efficient way of combining information across data stream so that detection is possible under the smallest detectable change magnitude. The challenge comes from the sparsity of change-points when only a small fraction of data streams undergo change at any point in time. The most successful approach to the sparsity issue so far has been the application of hard thresholding such that only local scores from data streams exhibiting significant changes are considered and added. However the identification of an optimal threshold is a difficult one. In particular it is unlikely that the same threshold is optimal for different levels of sparsity. We propose here a sparse likelihood score for identifying a sparse signal. The score is a likelihood ratio for testing between the null hypothesis of no change against an alternative hypothesis in which the change-points or signals are barely detectable. By the Neyman-Pearson Lemma this score has maximum detection power at the given alternative. The outcome is that we have a scoring of data streams that is successful in detecting at the boundary of the detectable region of signals and change-points. The likelihood score can be seen as a soft thresholding approach to sparse signal and change-point detection in which local scores that indicate small changes are down-weighted much more than local scores indicating large changes. We are able to show second-order optimality of the sparsity likelihood score in the sense of achieving successful detection at the minimum detectable order of change magnitude as well as at the minimum detection asymptotic constant with respect this order of change.

math.ST↗

arXiv2020

Infinite Arms Bandit: Optimality via Confidence Bounds

Berry et al. (1997) initiated the development of the infinite arms bandit problem. They derived a regret lower bound of all allocation strategies for Bernoulli rewards with uniform priors, and proposed strategies based on success runs. Bonald and Proutière (2013) proposed a two-target algorithm that achieves the regret lower bound, and extended optimality to Bernoulli rewards with general priors. We present here a confidence bound target (CBT) algorithm that achieves optimality for rewards that are bounded above. For each arm we construct a confidence bound and compare it against each other and a target value to determine if the arm should be sampled further. The target value depends on the assumed priors of the arm means. In the absence of information on the prior, the target value is determined empirically. Numerical studies here show that CBT is versatile and outperforms its competitors.

stat.ML↗

arXiv2019

Optimal UCB Adjustments for Large Arm Sizes

The regret lower bound of Lai and Robbins (1985), the gold standard for checking optimality of bandit algorithms, considers arm size fixed as sample size goes to infinity. We show that when arm size increases polynomially with sample size, a surprisingly smaller lower bound is achievable. This is because the larger experimentation costs when there are more arms permit regret savings by exploiting the best performer more often. In particular we are able to construct a UCB-Large algorithm that adaptively exploits more when there are more arms. It achieves the smaller lower bound and is thus optimal. Numerical experiments show that UCB-Large performs better than classical UCB that does not correct for arm size, and better than Thompson sampling.

math.ST↗

arXiv2019

The Multi-Armed Bandit Problem: An Efficient Non-Parametric Solution

Lai and Robbins (1985) and Lai (1987) provided efficient parametric solutions to the multi-armed bandit problem, showing that arm allocation via upper confidence bounds (UCB) achieves minimum regret. These bounds are constructed from the Kullback-Leibler information of the reward distributions, estimated from specified parametric families. In recent years there has been renewed interest in the multi-armed bandit problem due to new applications in machine learning algorithms and data analytics. Non-parametric arm allocation procedures like $ε$-greedy, Boltzmann exploration and BESA were studied, and modified versions of the UCB procedure were also analyzed under non-parametric settings. However unlike UCB these non-parametric procedures are not efficient under general parametric settings. In this paper we propose efficient non-parametric procedures.

math.ST↗

arXiv2016

Optimal sequential detection in multi-stream data

Consider a large number of detectors each generating a data stream. The task is to detect online, distribution changes in a small fraction of the data streams. Previous approaches to this problem include the use of mixture likelihood ratios and sum of CUSUMs. We provide here extensions and modifications of these approaches that are optimal in detecting normal mean shifts. We show how the (optimal) detection delay depends on the fraction of data streams undergoing distribution changes as the number of detectors goes to infinity. There are three detection domains. In the first domain for moderately large fractions, immediate detection is possible. In the second domain for smaller fractions, the detection delay grows logarithmically with the number of detectors, with an asymptotic constant extending those in sparse normal mixture detection. In the third domain for even smaller fractions, the detection delay lies in the framework of the classical detection delay formula of Lorden. We show that the optimal detection delay is achieved by the sum of detectability score transformations of either the partial scores or CUSUM scores of the data streams.

math.ST↗

arXiv2015

Multiscale Adaptive Inference on Conditional Moment Inequalities

This paper considers inference for conditional moment inequality models using a multiscale statistic. We derive the asymptotic distribution of this test statistic and use the result to propose feasible critical values that have a simple analytic formula, and to prove the asymptotic validity of a modified bootstrap procedure. The asymptotic distribution is extreme value, and the proof uses new techniques to overcome several technical obstacles. The test detects local alternatives that approach the identified set at the best rate among available tests in a broad class of models, and is adaptive to the smoothness properties of the data generating process. Our results also have implications for the use of moment selection procedures in this setting. We provide a monte carlo study and an empirical illustration to inference in a regression model with endogenously censored and missing data.

stat.AP↗

arXiv2015

Optimal detection of multi-sample aligned sparse signals

We describe, in the detection of multi-sample aligned sparse signals, the critical boundary separating detectable from nondetectable signals, and construct tests that achieve optimal detectability: penalized versions of the Berk-Jones and the higher-criticism test statistics evaluated over pooled scans, and an average likelihood ratio over the critical boundary. We show in our results an inter-play between the scale of the sequence length to signal length ratio, and the sparseness of the signals. In particular the difficulty of the detection problem is not noticeably affected unless this ratio grows exponentially with the number of sequences. We also recover the multiscale and sparse mixture testing problems as illustrative special cases.

math.ST↗

arXiv2014

A Sharp First Order Analysis of Feynman-Kac Particle Models

This article provides a new theory for the analysis of forward and backward particle approximations of Feynman-Kac models. Such formulae are found in a wide variety of applications and their numerical (particle) approximation are required due to their intractability. Under mild assumptions, we provide sharp and non-asymptotic first order expansions of these particle methods, potentially on path space and for possibly unbounded functions. These expansions allows one to consider upper and lower bound bias type estimates for a given time horizon $n$ and particle number $N$; these non-asymptotic estimates are of order $\mathcal{O}(n/N)$. Our approach is extended to tensor products of particle density profiles, leading to new sharp and non-asymptotic propagation of chaos estimates. The resulting upper and lower bound propagation of chaos estimates seems to be the first result of this kind for mean field particle models. As a by-product of our results, we also provide some analysis of the particle Gibbs sampler, providing first order expansions of the kernel and minorization estimates.

math.ST↗

arXiv2014

Theory of Parallel Particle Filters for Hidden Markov Models

The objective of this article is to study the asymptotic behavior of a new particle filtering approach in the context of hidden Markov models (HMMs). In particular, we develop an algorithm where the latent-state sequence is segmented into multiple shorter portions, with an estimation technique based upon a separate particle filter in each portion. The partitioning facilitates the use of parallel processing. Based upon this approach, we introduce new estimators of the latent states and likelihood which have similar or better variance properties compared to estimators derived from standard particle filters. Moreover due to parallelization there is less wall-clock computational time. We show that the likelihood function estimator is unbiased, prove central limit theorem convergences of estimators, and provide consistent in-sample estimation of the asymptotic variances. The theoretical analyses, supported by a numerical study, show that the segmentation reduces the variances in smoothed latent-state estimators, in addition to the savings in wall-clock time.

math.ST↗

arXiv2014

Detection with the scan and the average likelihood ratio

We investigate the performance of the scan (maximum likelihood ratio statistic) and of the average likelihood ratio statistic in the problem of detecting a deterministic signal with unknown spatial extent in the prototypical univariate sampled data model with white Gaussian noise. Our results show that the scan statistic, a popular tool for detection problems, is optimal only for the detection of signals with the smallest spatial extent. For signals with larger spatial extent the scan is suboptimal, and the power loss can be considerable. In contrast, the average likelihood ratio statistic is optimal for the detection of signals on all scales except the smallest ones, where its performance is only slightly suboptimal. We give rigorous mathematical statements of these results as well as heuristic explanations which suggest that the essence of these findings applies to detection problems quite generally, such as the detection of clusters in models involving densities or intensities or the detection of multivariate signals. We present a modification of the average likelihood ratio that yields optimal detection of signals with arbitrary spatial extent and which has the additional benefit of allowing for a fast computation of the statistic. In contrast, optimal detection with the scan seems to require the use of scale-dependent critical values.

stat.ME↗

arXiv2013

A general theory of particle filters in hidden Markov models and some applications

By making use of martingale representations, we derive the asymptotic normality of particle filters in hidden Markov models and a relatively simple formula for their asymptotic variances. Although repeated resamplings result in complicated dependence among the sample paths, the asymptotic variance formula and martingale representations lead to consistent estimates of the standard errors of the particle filter estimates of the hidden states.

math.ST↗

arXiv2012

A sequential Monte Carlo approach to computing tail probabilities in stochastic models

Sequential Monte Carlo methods which involve sequential importance sampling and resampling are shown to provide a versatile approach to computing probabilities of rare events. By making use of martingale representations of the sequential Monte Carlo estimators, we show how resampling weights can be chosen to yield logarithmically efficient Monte Carlo estimates of large deviation probabilities for multidimensional Markov random walks.

math.PR↗

arXiv2009

Detection of spatial clustering with average likelihood ratio test statistics

Generalized likelihood ratio (GLR) test statistics are often used in the detection of spatial clustering in case-control and case-population datasets to check for a significantly large proportion of cases within some scanning window. The traditional spatial scan test statistic takes the supremum GLR value over all windows, whereas the average likelihood ratio (ALR) test statistic that we consider here takes an average of the GLR values. Numerical experiments in the literature and in this paper show that the ALR test statistic has more power compared to the spatial scan statistic. We develop in this paper accurate tail probability approximations of the ALR test statistic that allow us to by-pass computer intensive Monte Carlo procedures to estimate $p$-values. In models that adjust for covariates, these Monte Carlo evaluations require an initial fitting of parameters that can result in very biased $p$-value estimates.

math.ST↗

arXiv2008

Importance Sampling of Word Patterns in DNA and Protein Sequences

Monte Carlo methods can provide accurate p-value estimates of word counting test statistics and are easy to implement. They are especially attractive when an asymptotic theory is absent or when either the search sequence or the word pattern is too short for the application of asymptotic formulae. Naive direct Monte Carlo is undesirable for the estimation of small probabilities because the associated rare events of interest are seldom generated. We propose instead efficient importance sampling algorithms that use controlled insertion of the desired word patterns on randomly generated sequences. The implementation is illustrated on word patterns of biological interest: Palindromes and inverted repeats, patterns arising from position specific weight matrices and co-occurrences of pairs of motifs.

stat.AP↗

arXiv2007

Maxima of Moving Sums in a Poisson Random Field

The extremal tail probabilities of moving sums in a marked Poisson random field is examined here. These sums are computed by adding up the weighted occurrences of events lying within a scanning set of fixed shape and size. Change of measure and analysis of local random fields are used to provide tail probabilities. The asymptotic constants are initially expressed in a form that seems hard to evaluate and do not seem to provide any additional information on the properties of the constants. A more sophisticated approach is then undertaken giving rise to an expression that is not only neater but also able to provide computable bounds. The technique used to obtain this constant can also be modified to work on continuous processes.

math.PR↗

arXiv2007

Efficient importance sampling for Monte Carlo evaluation of exceedance probabilities

Large deviation theory has provided important clues for the choice of importance sampling measures for Monte Carlo evaluation of exceedance probabilities. However, Glasserman and Wang [Ann. Appl. Probab. 7 (1997) 731--746] have given examples in which importance sampling measures that are consistent with large deviations can perform much worse than direct Monte Carlo. We address this problem by using certain mixtures of exponentially twisted measures for importance sampling. Their asymptotic optimality is established by using a new class of likelihood ratio martingales and renewal theory.

math.PR↗