DOCSLIB.ORG

Nonparametric Estimation of the Likelihood Ratio and Divergence Functionals

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Nonparametric Estimation of the Likelihood Ratio and Divergence Functionals 1 Nonparametric estimation of the likelihood ratio and divergence functionals 1 1 2 1 2 XuanLong Nguyen , Martin J. Wainwright ; and Michael I. Jordan ; 1 Department of Electrical Engineering and Computer Science 2 Department of Statistics University of California, Berkeley fxuanlong,wainwrig,[email protected] Abstract— We develop and analyze a nonparametric totic one; for example, the KL divergence emerges as the method for estimating the class of f-divergence functionals, asymptotic rate of the probability of error in Neyman- and the density ratio of two probability distributions. Pearson binary hypothesis testing (a result known as Our method is based on a non-asymptotic variational characterization of the f-divergence, which allows us to Stein's lemma). But it is also possible to provide non- cast the problem of estimating divergences in terms of asymptotic characterizations of the divergences; in par- risk minimization. We thus obtain an M-estimator for ticular, Fano's lemma shows that KL divergence provides divergences, based on a convex and differentiable opti- a lower bound on the error probability for decoding mization problem that can be solved efficiently We analyze error [5]. This paper is motivated by a non-asymptotic the consistency and convergence rates for this M-estimator given conditions only on the ratio of densities. characterization of f-divergence in the spirit of Fano's lemma, first explicated in our earlier work [14]. This characterization states that that there is an one-to-one I. INTRODUCTION correspondence between the family of f-divergences and Given samples from two (multivariate) probability the family of “surrogate loss functions”, such that the distributions P and Q, it is frequently of interest to (optimum) Bayes risk is equal to the negative of the di- estimate the values of functionals measuring the diver- vergence. In other words, any negative f-divergence can gence between the unknown P and Q. Of particular serve as a lower bound of a risk minimization problem. interest is the Kullback-Leibler (KL) divergence, but This variational characterization of divergence, stated the approach of this paper applies to the more general formally in Lemma 1, allows us estimate a divergence class of Ali-Silvey or f-divergences [1], [6]. An f- DÁ(P; Q) by solving a binary decision problem. Not divergence, to be defined formally in the sequel, is of the surprisingly, we show how the problem of estimating f- form DÁ(P; Q) = Á(dQ=dP)dP, where Á is a convex divergence is intrinsically linked to that of estimating the function of the likelihoodR ratio. likelihood ratio g0 = dP=dQ. Overall, we obtain an M- These divergences play fundamental roles in statis- estimator, whose optimal value estimates the divergence tics and information theory. In particular, divergences and optimizing argument estimates the likelihood ratio. are often used as measures of discrimination in binary Our estimator is nonparametric, in that it imposes no hypothesis testing and classification applications. Ex- strong assumptions on the form of the densities for P and amples include signal selection [11] and decentralized Q. We establish consistency of this estimator by exploit- detection [14], where f-divergences are used to solve ing analysis techniques for M-estimators in the setting experimental design problems. The Shannon mutual of nonparametric density estimation and regression [18], information (a particular type of KL divergence), in [20]. At a high level, the key to the proof is suitable addition to its role in coding theorems, is often used as a control on the modulus of continuity of the suprema measure of independence to be extremized in dimension- of two empirical processes, one for each of P and Q, ality reduction and feature selection. In all of these cases, with respect to a metric defined over density ratios. if divergences are to be used as objective functionals, one This metric turns out to be a surrogate lower bound has to be able to estimate them efficiently from data. of a Bregman divergence defined on a pair of density There are two ways in which divergences are typically ratios. In this way, we not only establish consistency of characterized. The classical characterization is an asymp- our estimator, but also obtain convergence rates. As one concrete example, when the likehood ratio g0 lies in a The KL divergence is a special case of a broader function class G of smoothness ® with ® > d=2, where class of divergences known as Ali-Silvey distances, or d is the number of data dimensions, our estimator of f-divergences [6], [1]: the likelihood ratio achieves the optimal minimax rate ¡®=(2®+d) n according to the Hellinger metric, while the DÁ(P; Q) = p0Á(q0=p0) d¹; Z divergence estimator also achieves the rate n¡®=(2®+d). In abstract terms, an f-divergence can be viewed as an where Á : R ! R¹ is a convex function. Different choices integral functional of a pair of densities. While there are of Á result in many divergences that play important relatively little work focusing on integral functionals for roles in information theory and statistics, including the pairs of densities (such as the f-divergences of interest variational distance, Hellinger distance, KL divergence here), there is an extensive literature on the estimation of and so on (see, e.g., [17]). an integral functionals of the form Á(p)p, where p the Since Á is a convex function, by Legendre-Fenchel density of an unknown probability distribR ution. Work on convex duality [16] we can write: this topic dates back to the 1970s [9], [13]; see also [2], Á(u) = sup(uv ¡ Á¤(v)); [3], [12] and the references therein. There are also a v2R number of papers that focus specifically on the entropy ¤ functional (see, e.g., [7], [10], [8]). where Á is the convex conjugate of Á. As a result, In a separate line of work, Wang et al. [21] proposed ¤ DÁ(P; Q) = p0 sup(fq0=p0 ¡ Á (f)) d¹ a histogram-based KL estimator, which is based on the Z f estimation of the likelihood ratio by building partitions of equivalent (empirical) Q-measure. The estimator was = sup f dQ ¡ Á¤(f) dP ; µ ¶ empirically shown to outperform direct plug-in methods, f Z Z but no theoretical convergence rate analysis was given. A where the supremum is taken over all measurable func- concern with histogram-based methods are their possible tions f : X ! R, and f dP denotes the expectation inefficiency, in both statistical and computational terms, of f under distribution PR. It is easy to see that equality when applied to higher dimensional data. Our prelim- in the supremum is attained for functions f such that ¤ inary empirical results [15] suggest that our estimator q0=p0 2 @Á (f), where q0; p0 and f are evaluated at any exhibits comparable or superior convergence rates in a x 2 X . By convex duality, this is true if f 2 @Á(q0=p0) number of examples. for any x 2 X . Thus, we have proved the following: The remainder of this paper is organized as follows. In Lemma 1. Letting be any function class in R, Section II, we describe a general variational character- F X ! there holds: ization of f-divergence, and derive an M-estimator for the KL divergence and the likelihood ratio. Section III is ¤ DÁ(P; Q) ¸ sup f dQ ¡ Á (f) dP: (1) devoted to the analysis of consistency and convergence f2F Z rates. In Section IV we briefly discuss how our analysis Furthermore, equality holds if F \ @Á(q =p ) 6= ;. extends to general f-divergences. Additional results and 0 0 complete proofs of all theorems can be be found in [15]. B. An M-estimator of density ratio and KL divergence II. M -ESTIMATOR FORMULATION Returning to the KL divergence, Á has the form We begin by describing an M-estimator formulation Á(u) = ¡ log(u) for u > 0 and +1 for u · 0. for KL divergence and the density ratio. ¤ The convex dual of Á is Á (v) = supu(uv ¡ Á(u)) = A. Variational characterization of f-divergence ¡1¡log(¡v) if u < 0 and +1 otherwise. By Lemma 1, Let X1; : : : ; Xn be n i.i.d. random variables drawn from an unknown distribution P; similarly, let Y ; : : : ; Y DK (P; Q) = sup f dQ ¡ (¡1 ¡ log(¡f)) dP 1 n f<0 Z Z be n random variables drawn from an unknown distribu- tion Q. Assume that both are absolutely continuous with = sup log g dP ¡ gdQ + 1: (2) g>0 Z Z respect to Lebesgue measure ¹, with positive densities p0 d and q0, respectively, on some compact domain X ½ R . In addition, the supremum is attained at g0 = p0=q0. This The KL divergence between P and Q is defined as: motivates the following estimator of the KL divergence: Let G be a function class of X ! R+, and dPn and DK (P; Q) = p0 log(p0=q0) d¹: Z dQn denote the expectation under empiricalR measures R 2 Pn and Qn, respectively, and consider the following rates. We bound E1(G) in terms of the following empir- optimization problem: ical processes: g D^ = sup log g dP ¡ gdQ + 1: (3) vn(G) = sup log d(Pn¡P)¡ (g¡g0)d(Qn¡Q) : K n n ¯ Z g Z ¯ g2G Z Z g2G¯ 0 ¯ ¯ ¯ In practice we generally choose G to be a convex ¯ ¯ wn(g0) = log g0 d(Pn ¡ P) ¡ g0d(Qn ¡ Q) : function class, which turns (3) into a convex optimization ¯ Z ¯ ¯ ¯ problem that can be solved efficiently [15]. Suppose ¯ ¯ Note that by ¯construction, we have: ¯ that the supremum is attained at g^n.
Load more

Recommended publications
  • IR-IITBHU at TREC 2016 Open Search Track: Retrieving Documents Using Divergence from Randomness Model in Terrier
    IR-IITBHU at TREC 2016 Open Search Track: Retrieving documents using Divergence From Randomness model in Terrier Mitodru Niyogi1 and Sukomal Pal2 1Department of Information Technology, Government College of Engineering & Ceramic Technology, Kolkata 2Department of Computer Science & Engineering, Indian Institute of Technology(BHU), Varanasi Abstract In our participation at TREC 2016 Open Search Track which focuses on ad-hoc scientic literature search, we used Terrier, a modular and a scalable Information Retrieval framework as a tool to rank documents. The organizers provided live data as documents, queries and user interac- tions from real search engine that were available through Living Lab API framework. The data was then converted into TREC format to be used in Terrier. We used Divergence from Randomness (DFR) model, specically, the Inverse expected document frequency model for randomness, the ratio of two Bernoulli's processes for rst normalisation, and normalisation 2 for term frequency normalization with natural logarithm, i.e., In_expC2 model to rank the available documents in response to a set of queries. Al- together we submit 391 runs for sites CiteSeerX and SSOAR to the Open Search Track via the Living Lab API framework. We received an `out- come' of 0.72 for test queries and 0.62 for train queries of site CiteSeerX at the end of Round 3 Test Period where, the outcome is computed as: #wins / (#wins + #losses). A `win' occurs when the participant achieves more clicks on his documents than those of the site and `loss' otherwise. Complete relevance judgments is awaited at the moment. We look forward to getting the users' feedback and work further with them.
    [Show full text]
  • Subnational Inequality Divergence
    Subnational Inequality Divergence Tom VanHeuvelen1 University of Minnesota Department of Sociology Abstract How have inequality levels across local labor markets in the subnational United States changed over the past eight decades? In this study, I examine inequality divergence, or the inequality of inequalities. While divergence trends of central tendencies such as per capita income have been well documented, less is known about the descriptive trends or contributing mechanisms for inequality. In this study, I construct wage inequality measures in 722 local labor markets covering the entire contiguous United States across 22 waves of Census and American Community Survey data from 1940-2019 to assess the historical trends of inequality divergence. I apply variance decomposition and counterfactual techniques to develop main conclusions. Inequality divergence follows a u-shaped pattern, declining through 1990 but with contemporary divergence at as high a level as any time in the past 80 years. Early era convergence occurred broadly and primarily worked to reduce interregional differences, whereas modern inequality divergence operates through a combination of novel mechanisms, most notably through highly unequal urban areas separating from other labor markets. Overall, results show geographical fragmentation of inequality underneath overall inequality growth in recent years, highlighting the fundamental importance of spatial trends for broader stratification outcomes. 1 Correspondence: [email protected]. A previous version of this manuscript was presented at the 2021 Population Association of American annual conference. Thank you to Jane VanHeuvelen and Peter Catron for their helpful comments. Recent changes in the United States have situated geographical residence as a key pillar of the unequal distribution of economic resources (Austin et al.
    [Show full text]
  • A Generalized Divergence for Statistical Inference
    A GENERALIZED DIVERGENCE FOR STATISTICAL INFERENCE Abhik Ghosh, Ian R. Harris, Avijit Maji Ayanendranath Basu and Leandro Pardo TECHNICAL REPORT NO. BIRU/2013/3 2013 BAYESIAN AND INTERDISCIPLINARY RESEARCH UNIT INDIAN STATISTICAL INSTITUTE 203, Barrackpore Trunk Road Kolkata – 700 108 INDIA A Generalized Divergence for Statistical Inference Abhik Ghosh Indian Statistical Institute, Kolkata, India. Ian R. Harris Southern Methodist University, Dallas, USA. Avijit Maji Indian Statistical Institute, Kolkata, India. Ayanendranath Basu Indian Statistical Institute, Kolkata, India. Leandro Pardo Complutense University, Madrid, Spain. Summary. The power divergence (PD) and the density power divergence (DPD) families have proved to be useful tools in the area of robust inference. The families have striking similarities, but also have fundamental differences; yet both families are extremely useful in their own ways. In this paper we provide a comprehensive description of the two families and tie in their role in statistical theory and practice. At the end, the families are seen to be a part of a superfamily which contains both of these families as special cases. In the process, the limitation of the influence function as an effective descriptor of the robustness of the estimator is also demonstrated. Keywords: Robust Estimation, Divergence, Influence Function 2 Ghosh et al. 1. Introduction The density-based minimum divergence approach is an useful technique in para- metric inference. Here the closeness of the data and the model is quantified by a suitable measure of density-based divergence between the data density and the model density. Many of these methods have been particularly useful because of the strong robustness properties that they inherently possess.
    [Show full text]
  • Divergence Measures for Statistical Data Processing—An Annotated Bibliography
    Signal Processing 93 (2013) 621–633 Contents lists available at SciVerse ScienceDirect Signal Processing journal homepage: www.elsevier.com/locate/sigpro Review Divergence measures for statistical data processing—An annotated bibliography Michele Basseville n,1 IRISA, Campus de Beaulieu, 35042 Rennes Cedex, France article info abstract Article history: This paper provides an annotated bibliography for investigations based on or related to Received 19 April 2012 divergence measures for statistical data processing and inference problems. Received in revised form & 2012 Elsevier B.V. All rights reserved. 17 July 2012 Accepted 5 September 2012 Available online 14 September 2012 Keywords: Divergence Distance Information f-Divergence Bregman divergence Learning Estimation Detection Classification Recognition Compression Indexing Contents 1. Introduction ......................................................................................... 621 2. f-Divergences ........................................................................................ 622 3. Bregman divergences.................................................................................. 623 4. a-Divergences ....................................................................................... 624 5. Handling more than two distributions .................................................................... 625 6. Inference based on entropy and divergence criteria.......................................................... 626 7. Spectral divergence measures . .......................................................................
    [Show full text]
  • Robust Hypothesis Testing with Α-Divergence Gokhan¨ Gul,¨ Student Member, IEEE, Abdelhak M
    1 Robust Hypothesis Testing with α-Divergence Gokhan¨ Gul,¨ Student Member, IEEE, Abdelhak M. Zoubir, Fellow, IEEE, Abstract—A robust minimax test for two composite hypotheses, effects that go unmodeled [2]. which are determined by the neighborhoods of two nominal dis- Imprecise knowledge of F0 or F1 leads, in general, to per- tributions with respect to a set of distances - called α−divergence formance degradation and a useful approach is to extend the distances, is proposed. Sion’s minimax theorem is adopted to characterize the saddle value condition. Least favorable distri- known model by accepting a set of distributions Fi, under each butions, the robust decision rule and the robust likelihood ratio hypothesis Hi, that are populated by probability distributions test are derived. If the nominal probability distributions satisfy Gi, which are at the neighborhood of the nominal distribution a symmetry condition, the design procedure is shown to be Fi based on some distance D [1]. Under some mild conditions simplified considerably. The parameters controlling the degree of on D, it can be shown that the best (error minimizing) decision robustness are bounded from above and the bounds are shown to ^ be resulting from a solution of a set of equations. The simulations rule δ for the worst case (error maximizing) pair of probability ^ ^ performed evaluate and exemplify the theoretical derivations. distributions (G0; G1) 2 F0 × F1 accepts a saddle value. Therefore, such a test design guarantees a certain level of Index Terms—Detection, hypothesis testing, robustness, least favorable distributions, minimax optimization, likelihood ratio detection at all times. This type of optimization is known test.
    [Show full text]
  • Divergence Tendencies in the European Integration Process: a Danger for the Sustainability of the E(M)U?
    Journal of Risk and Financial Management Article Divergence Tendencies in the European Integration Process: A Danger for the Sustainability of the E(M)U? Linda Glawe and Helmut Wagner * Faculty of Economics, University of Hagen, 58084 Hagen, Germany; [email protected] * Correspondence: [email protected] Abstract: The European integration process started with the aim of reducing the differences in income and/or living standards between the participating countries over time. To achieve this, a certain alignment of institutions and structures was seen as a necessary precondition. While the goal of this income and institutional convergence was successfully achieved over a long period of time, this convergence development has weakened or even turned into divergence in the last one to two decades. This paper provides an overview of the empirical evidence for these convergence and divergence developments and develops policy implications (the challenges and possible ways out). Keywords: European integration process; convergence dynamics; divergence; institutional quality; economic growth and economic development; European Union JEL Classification: F55; O10; O11; O43; O52 Citation: Glawe, Linda, and Helmut Wagner. 2021. Divergence Tendencies 1. Introduction in the European Integration Process: It is conventional wisdom that the larger and more complex a project is, the greater the A Danger for the Sustainability of the risk that it will fail. This also applies to the project of European integration. This project was E(M)U? Journal of Risk and Financial initially small and therefore manageable and also successful. Over time, however, more and Management 14: 104. https:// more extensions were made (in terms of scope, and in terms of areas of application/depth), doi.org/10.3390/jrfm14030104 so that it has become increasingly difficult and complex to keep the situation/project stable/sustainable.
    [Show full text]
  • GDP PER CAPITA VERSUS MEDIAN HOUSEHOLD INCOME: WHAT GIVES RISE to DIVERGENCE OVER TIME? Brian Nolan, Max Roser, and Stefan
    GDP PER CAPITA VERSUS MEDIAN HOUSEHOLD INCOME: WHAT GIVES RISE TO DIVERGENCE OVER TIME? Brian Nolan, Max Roser, and Stefan Thewissen INET Oxford WorkinG Paper no. 2016-03 Employment, Equity and Growth proGramme GDP per capita versus median household income: What gives rise to divergence over time?⇤ § Brian Nolan,† Max Roser,‡ and Stefan Thewissen May 2016 Abstract Divergence between the evolution of GDP per capita and the income of a‘typical’householdasmeasuredinhouseholdsurveysisgivingrisetoa range of serious concerns, especially in the USA. This paper investigates the extent of that divergence and the factors that contribute to it across 27 OECD countries, using data from OECD National Accounts and the Luxembourg Income Study. While GDP per capita has risen faster than median household income in most of these countries over the period these data cover, the size of that divergence varied very substantially, with the USA a clear outlier. The paper distinguishes a number of factors contributing to such a divergence, and finds wide variation across countries in the impact of the various factors. Further, both the extent of that divergence and the role of the various contributory factors vary widely over time for most of the countries studied. These findings have serious implications for the monitoring and assessment of changes in household incomes and living standards over time. keywords: Economic growth, inequality, household incomes ⇤The authors are grateful for comments and advice from Tony Atkinson (University of Oxford) and Facundo Alvaredo (Paris School of Economics). †Institute for New Economic Thinking at the Oxford Martin School, Department of Social Policy and Intervention, and Nuffield College, University of Oxford.
    [Show full text]
  • On the Use of Entropy to Improve Model Selection Criteria
    entropy Article On the Use of Entropy to Improve Model Selection Criteria Andrea Murari 1, Emmanuele Peluso 2, Francesco Cianfrani 2, Pasquale Gaudio 2 and Michele Lungaroni 2,* 1 Consorzio RFX (CNR, ENEA, INFN, Universita’ di Padova, Acciaierie Venete SpA), 35127 Padova, Italy; [email protected] 2 Department of Industrial Engineering, University of Rome “Tor Vergata”, 00133 Roma, Italy; [email protected] (E.P.); [email protected] (F.C.); [email protected] (P.G.) * Correspondence: [email protected]; Tel.: +39-(0)6-7259-7196 Received: 21 January 2019; Accepted: 10 April 2019; Published: 12 April 2019 Abstract: The most widely used forms of model selection criteria, the Bayesian Information Criterion (BIC) and the Akaike Information Criterion (AIC), are expressed in terms of synthetic indicators of the residual distribution: the variance and the mean-squared error of the residuals respectively. In many applications in science, the noise affecting the data can be expected to have a Gaussian distribution. Therefore, at the same level of variance and mean-squared error, models, whose residuals are more uniformly distributed, should be favoured. The degree of uniformity of the residuals can be quantified by the Shannon entropy. Including the Shannon entropy in the BIC and AIC expressions improves significantly these criteria. The better performances have been demonstrated empirically with a series of simulations for various classes of functions and for different levels and statistics of the noise. In presence of outliers, a better treatment of the errors, using the Geodesic Distance, has proved essential. Keywords: Model Selection Criteria; Bayesian Information Criterion (BIC); Akaike Information Criterion (AIC); Shannon Entropy; Geodesic Distance 1.
    [Show full text]
  • Digitizing Omics Profiles by Divergence from a Baseline
    Digitizing omics profiles by divergence from a baseline INAUGURAL ARTICLE Wikum Dinalankaraa,1, Qian Keb,1, Yiran Xub, Lanlan Jib, Nicole Paganea, Anching Liena, Tejasvi Matama, Elana J. Fertiga, Nathan D. Pricec, Laurent Younesb, Luigi Marchionnia,2, and Donald Gemanb,2 aDepartment of Oncology, Johns Hopkins University School of Medicine, Baltimore, MD 21205; bDepartment of Applied Mathematics and Statistics, Johns Hopkins University, Baltimore, MD 21218; and cInstitute for Systems Biology, Seattle, WA 98109 This contribution is part of the special series of Inaugural Articles by members of the National Academy of Sciences elected in 2015. Contributed by Donald Geman, March 19, 2018 (sent for review December 13, 2017; reviewed by Nicholas P. Tatonetti and Bin Yu) Data collected from omics technologies have revealed perva- relative to a baseline, may be essential to fully realize the poten- sive heterogeneity and stochasticity of molecular states within tial of high-dimensional omics assays; in principle, they pro- and between phenotypes. A prominent example of such het- vide unprecedented quantification of normal and disease-specific erogeneity occurs between genome-wide mRNA, microRNA, and variation, which can inform clinical approaches for diagnosis and methylation profiles from one individual tumor to another, even prevention of complex diseases. within a cancer subtype. However, current methods in bioinfor- To develop this high-dimensional analogue to low-dimensional matics, such as detecting differentially expressed genes or CpG clinical tests, we present a unified framework to characterize sites, are population-based and therefore do not effectively model normal and nonnormal variation by quantizing high-resolution intersample diversity. Here we introduce a unified theory to quan- measurements into a few discrete states based on divergence tify sample-level heterogeneity that is applicable to a single omics from a baseline.
    [Show full text]
  • Stagnating Median Incomes Despite Economic Growth: Explaining the Divergence in 27 OECD Countries
    12/3/2018 Economic growth with stagnating median incomes: New analysis | VOX, CEPR Policy Portal VOX CEPR Policy Portal Create account | Login | Subscribe Research-based policy analysis and commentary from leading economists Search Columns Video Vox Vox Talks Publications Blogs&Reviews People Debates Events About By Topic By Date By Reads By Tag Stagnating median incomes despite economic growth: Explaining the divergence in 27 OECD countries Brian Nolan, Max Roser, Stefan Thewissen 27 August 2016 Brian Nolan With inequality rising and household incomes across developed countries stagnating, accurate monitoring of Professor of Social Policy and living standards cannot be achieved by relying on GDP per capita alone. This column analyses the path of Director of Employment, Equity, divergence between household income and GDP per capita for 27 OECD countries. It finds several reasons why and Growth Programme, Institute GDP per capita has outpaced median incomes, and recommends assigning median income a central place in for New Economic Thinking; official monitoring and assessment of living standards over time. Senior Research Fellow, Nuffield College, University of Oxford 255 A A Increasing inequality and stagnating incomes for Related ordinary workers and their families are now at the centre Growth quality: A new index of economic and political debate across the rich Montfort Mlachila, René Tapsoba, Sampawende Tapsoba countries. Against that backdrop, the fact that median Max Roser household income has lagged so far behind growth in The turtle’s progress: Secular stagnation meets the headwinds Programme Director, Oxford GDP per capita in the US – widely attributed to growing Robert J. Gordon Martin School, University of inequality – has reinforced concerns about relying on Global income distribution since 1988 Oxford GDP as the core measure of economic prosperity Christoph Lakner, Branko Milanovic (Stiglitz et al.
    [Show full text]
  • Lecture 7: Hypothesis Testing and KL Divergence 1 Introducing The
    ECE 830 Spring 2015 Statistical Signal Processing instructor: R. Nowak Lecture 7: Hypothesis Testing and KL Divergence 1 Introducing the Kullback-Leibler Divergence iid Suppose X1;X2;:::;Xn ∼ q(x) and we have two models for q(x), p0(x) and p1(x). In past lectures we have seen that the likelihood ratio test (LRT) is optimal, assuming that q is p0 or p1. The error probabilities can be computed numerically in many cases. The error probabilities converge to 0 as the number of samples n grows, but numerical calculations do not always yield insight into rate of convergence. In this lecture we will see that the rate is exponential in n and parameterized the Kullback-Leibler (KL) divergence, which quantifies the differences between the distributions p0 and p1. Our analysis will also give insight into the performance of the LRT when q is neither p0 nor p1. This is important since in practice p0 and p1 may be imperfect models for reality, q in this context. The LRT acts as one would expect in such cases, it picks the model that is closest (in the sense of KL divergence) to q. To begin our discusion, recall the likelihood ratio is n Y p1(xi) Λ = p (x ) i=1 0 i The log likelihood ratio, normalized by dividing by n, is then n 1 X p1(xi) Λ^ = log n n p (x ) i=1 0 i p1(xi) Note that Λ^ n is itself a random variable, and is in fact a sum of iid random variables Li = log which p0(xi) are independent because the xi are.
    [Show full text]
  • Improved Security Proofs in Lattice-Based Cryptography: Using the Rényi Divergence Rather Than the Statistical Distance
    Improved security proofs in lattice-based cryptography: using the Rényi divergence rather than the statistical distance Shi Bai1, Tancrède Lepoint3, Adeline Roux-Langlois4, Amin Sakzad5, Damien Stehlé2, and Ron Steinfeld5 1 Department of Mathematical Sciences, Florida Atlantic University, USA. [email protected] – http://cosweb1.fau.edu/~sbai 2 ENS de Lyon, Laboratoire LIP (U. Lyon, CNRS, ENSL, INRIA, UCBL), France. [email protected] – http://perso.ens-lyon.fr/damien.stehle 3 SRI International, USA. [email protected] – https://tlepoint.github.io/ 4 CNRS/IRISA, Rennes, France. [email protected] – http://people.irisa.fr/Adeline.Roux-Langlois/ 5 Faculty of Information Technology, Monash University, Clayton, Australia. [email protected], [email protected] – http://users.monash.edu.au/~rste/ Abstract. The Rényi divergence is a measure of closeness of two prob- ability distributions. We show that it can often be used as an alternative to the statistical distance in security proofs for lattice-based cryptogra- phy. Using the Rényi divergence is particularly suited for security proofs of primitives in which the attacker is required to solve a search problem (e.g., forging a signature). We show that it may also be used in the case of distinguishing problems (e.g., semantic security of encryption schemes), when they enjoy a public sampleability property. The techniques lead to security proofs for schemes with smaller parameters, and sometimes to simpler security proofs than the existing ones. 1 Introduction Let D1 and D2 be two non-vanishing probability distributions over a common measurable support X. Let a ∈ (1, +∞). The Rényi divergence [Rén61,vEH14] (RD for short) Ra(D1kD2) of order a between D1 and D2 is defined as the a−1 ((a − 1)th root of the) expected value of (D1(x)/D2(x)) over the randomness of x sampled from D1.
    [Show full text]