FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W2103606110> ?p ?o ?g. }

• W2103606110 abstract "The accurate evaluation of marginal likelihood integrals is a difficult fundamental problem in Bayesian inference that has important applications in machine learning and computational biology. Following the recent success of algebraic statistics in frequentist inference and inspired by Watanabe's foundational approach to singular learning theory, the goal of this dissertation is to study algebraic, geometric and combinatorial methods for computing Bayesian integrals effectively, and to explore the rich mathematical theories that arise in this connection between statistics and algebraic geometry. For these integrals, we investigate their exact evaluation for small samples and their asymptotics for large samples.According to Watanabe, the key to understanding singular models lies in desingularizing the Kullback-Leibler function K(w) of the model at the true distribution. This step puts the model in a standard form so that various central limit theorems can be applied. While general algorithms exist for desingularizing any analytic function, applying them to non-polynomial functions such as K(w) can be computationally expensive. Many singular models are however represented as regular models whose parameters are polynomial functions of new parameters. Discrete models and multivariate Gaussian models are all examples. We call them regularly parametrized models. One of our main contributions is showing how this polynomiality can be exploited by defining fiber ideals for singular models and relating the properties of these algebraic objects to the statistics. In particular, we prove that a model is put in standard form if we monomialize the corresponding fiber ideal. As a corollary, the learning coefficient of a model is equal to the real log canonical threshold (RLCT) of the fiber ideal.While complex log canonical thresholds are well-studied in algebraic geometry, little is known about their real analogs. In Chapter 4, we prove their fundamental properties and simple rules of computation. We also extend Varchenko's notion of Newton polyhedra and nondegeneracy for functions to ideals. Using these methods, we discover a formula for the RLCT of a monomial ideal with respect to a monomial amplitude. For all other ideals, this formula is an upper bound for their RLCT. Our tools are then applied to a difficult statistical example involving a naive Bayesian network with two ternary random variables.Because our statistical models are defined over compact semianalytic parameter spaces W, we need to extend standard asymptotic theory of real analytic functions over neighborhoods of the origin to functions over domains like W. Chapter 3 summarizes these results which are critical for other proofs in this dissertation. We also give explicit formulas for the full asymptotic expansion of a Laplace integral over W in terms of the Laurent coefficients of the associated zeta function. In Chapter 5, we apply these formulas to Laplace integrals Z(n) with nondegenerate phase functions, and describe algorithms for computing the coefficient C in the first term asymptotics Z(n) = C n^{-l} (log n)^{t-1}. Procedures for calculating all higher order coefficients are also developed and explained.Watanabe's treatment of singular models assumes knowledge of the true distribution. In this dissertation, we also explore marginal likelihood integrals of exponential families given data where the true distribution is unknown. This is the context in which Schwarz, Haughton, and Geiger and Rusakov studied the Bayesian Information Criterion (BIC). We find here that the log likelihood ratio of the data is equal to the Kullback-Leibler function of the model at the maximum likelihood distribution. Therefore, all the methods we developed for Kullback-Leibler functions apply, so we describe how to compute the full asymptotics of the marginal likelihood integral by monomializing the associated fiber ideal.Lastly, to complement developments in asymptotic estimation as well as in Markov Chain Monte Carlo (MCMC) estimation, we present, in Chapter 2, symbolic algorithms for computing marginal likelihood integrals exactly for discrete data of small samples. The underlying statistical models are mixtures of independent distributions, or, in geometric language, secant varieties of Segre-Veronese varieties. For these models, the numerical value of the integral is a rational number, and exact evaluation means computing that rational number rather than a floating point approximation. These exact results provide a gold standard with which approximation methods can be compared." @default.
• W2103606110 created "2016-06-24" @default.
• W2103606110 creator A5048857206 @default.
• W2103606110 date "2011-01-01" @default.
• W2103606110 modified "2023-09-28" @default.
• W2103606110 title "Algebraic methods for evaluating integrals In Bayesian statistics" @default.
• W2103606110 cites W128650179 @default.
• W2103606110 cites W1483218536 @default.
• W2103606110 cites W1549105346 @default.
• W2103606110 cites W1567331820 @default.
• W2103606110 cites W1574365947 @default.
• W2103606110 cites W1580842374 @default.
• W2103606110 cites W1584154100 @default.
• W2103606110 cites W1587445444 @default.
• W2103606110 cites W1595099852 @default.
• W2103606110 cites W1615427479 @default.
• W2103606110 cites W1653612104 @default.
• W2103606110 cites W1656717670 @default.
• W2103606110 cites W1658982721 @default.
• W2103606110 cites W173872297 @default.
• W2103606110 cites W1758916186 @default.
• W2103606110 cites W1771599666 @default.
• W2103606110 cites W1772827447 @default.
• W2103606110 cites W1856578837 @default.
• W2103606110 cites W1942683912 @default.
• W2103606110 cites W1968936725 @default.
• W2103606110 cites W1972789140 @default.
• W2103606110 cites W1978102570 @default.
• W2103606110 cites W1981599791 @default.
• W2103606110 cites W1987006424 @default.
• W2103606110 cites W1991644091 @default.
• W2103606110 cites W2010061190 @default.
• W2103606110 cites W2040177378 @default.
• W2103606110 cites W2049633694 @default.
• W2103606110 cites W2060161465 @default.
• W2103606110 cites W2064319051 @default.
• W2103606110 cites W2069641044 @default.
• W2103606110 cites W2072778147 @default.
• W2103606110 cites W2080574402 @default.
• W2103606110 cites W2086331397 @default.
• W2103606110 cites W2106568142 @default.
• W2103606110 cites W2110009816 @default.
• W2103606110 cites W2120217353 @default.
• W2103606110 cites W2120340025 @default.
• W2103606110 cites W2124430067 @default.
• W2103606110 cites W2146040013 @default.
• W2103606110 cites W2146929214 @default.
• W2103606110 cites W2149301289 @default.
• W2103606110 cites W2150126448 @default.
• W2103606110 cites W2157436031 @default.
• W2103606110 cites W2159626963 @default.
• W2103606110 cites W2168175751 @default.
• W2103606110 cites W2168359188 @default.
• W2103606110 cites W2468976128 @default.
• W2103606110 cites W2963363436 @default.
• W2103606110 cites W3099019541 @default.
• W2103606110 cites W3101303041 @default.
• W2103606110 cites W3101346901 @default.
• W2103606110 cites W3144205924 @default.
• W2103606110 cites W963196086 @default.
• W2103606110 hasPublicationYear "2011" @default.
• W2103606110 type Work @default.
• W2103606110 sameAs 2103606110 @default.
• W2103606110 citedByCount "6" @default.
• W2103606110 countsByYear W21036061102012 @default.
• W2103606110 countsByYear W21036061102013 @default.
• W2103606110 countsByYear W21036061102020 @default.
• W2103606110 countsByYear W21036061102021 @default.
• W2103606110 crossrefType "book" @default.
• W2103606110 hasAuthorship W2103606110A5048857206 @default.
• W2103606110 hasConcept C101112237 @default.
• W2103606110 hasConcept C105795698 @default.
• W2103606110 hasConcept C107673813 @default.
• W2103606110 hasConcept C134261354 @default.
• W2103606110 hasConcept C134306372 @default.
• W2103606110 hasConcept C136119220 @default.
• W2103606110 hasConcept C14036430 @default.
• W2103606110 hasConcept C148607811 @default.
• W2103606110 hasConcept C154945302 @default.
• W2103606110 hasConcept C160234255 @default.
• W2103606110 hasConcept C162376815 @default.
• W2103606110 hasConcept C186219872 @default.
• W2103606110 hasConcept C201482947 @default.
• W2103606110 hasConcept C202444582 @default.
• W2103606110 hasConcept C2776214188 @default.
• W2103606110 hasConcept C28826006 @default.
• W2103606110 hasConcept C33923547 @default.
• W2103606110 hasConcept C41008148 @default.
• W2103606110 hasConcept C51544822 @default.
• W2103606110 hasConcept C78045399 @default.
• W2103606110 hasConcept C78458016 @default.
• W2103606110 hasConcept C86803240 @default.
• W2103606110 hasConcept C90119067 @default.
• W2103606110 hasConcept C9376300 @default.
• W2103606110 hasConceptScore W2103606110C101112237 @default.
• W2103606110 hasConceptScore W2103606110C105795698 @default.
• W2103606110 hasConceptScore W2103606110C107673813 @default.
• W2103606110 hasConceptScore W2103606110C134261354 @default.
• W2103606110 hasConceptScore W2103606110C134306372 @default.
• W2103606110 hasConceptScore W2103606110C136119220 @default.

• next