Basics of Information Theory: Markku Juntti. Basic concepts and tools 1 Introduction 2 Entropy, relative entropy and mutual information

Similar documents
3. Basic Concepts: Consequences and Properties

Differential Entropy 吳家麟教授

. The set of these sums. be a partition of [ ab, ]. Consider the sum f( x) f( x 1)

Basics of Information Theory: Markku Juntti. Basic concepts and tools 1 Introduction 2 Entropy, relative entropy and mutual information

Lecture 3 Probability review (cont d)

Entropies & Information Theory

Module 7. Lecture 7: Statistical parameter estimation

Chapter 5 Properties of a Random Sample

ρ < 1 be five real numbers. The

5. Data Compression. Review of Last Lecture. Outline of the Lecture. Course Overview. Basics of Information Theory: Markku Juntti

( ) = ( ) ( ) Chapter 13 Asymptotic Theory and Stochastic Regressors. Stochastic regressors model

CHAPTER VI Statistical Analysis of Experimental Data

UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS

Continuous Random Variables: Conditioning, Expectation and Independence

= 2. Statistic - function that doesn't depend on any of the known parameters; examples:

Chapter 4 Multiple Random Variables

UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS

Functions of Random Variables

X ε ) = 0, or equivalently, lim

Summary of the lecture in Biostatistics

Introduction to Numerical Differentiation and Interpolation March 10, !=1 1!=1 2!=2 3!=6 4!=24 5!= 120

Special Instructions / Useful Data

Discrete Mathematics and Probability Theory Fall 2016 Seshia and Walrand DIS 10b

Unsupervised Learning and Other Neural Networks

Simulation Output Analysis

Chapter 14 Logistic Regression Models

4 Round-Off and Truncation Errors

LECTURE 8: Topics in Chaos Ricker Equation. Period doubling bifurcation. Period doubling cascade. A Quadratic Equation Ricker Equation 1.0. x x 4 0.

Lecture 7. Confidence Intervals and Hypothesis Tests in the Simple CLR Model

Idea is to sample from a different distribution that picks points in important regions of the sample space. Want ( ) ( ) ( ) E f X = f x g x dx

Outline. Point Pattern Analysis Part I. Revisit IRP/CSR

Lecture Notes Types of economic variables

ENGI 4421 Joint Probability Distributions Page Joint Probability Distributions [Navidi sections 2.5 and 2.6; Devore sections

Asymptotic Formulas Composite Numbers II

Random Variate Generation ENM 307 SIMULATION. Anadolu Üniversitesi, Endüstri Mühendisliği Bölümü. Yrd. Doç. Dr. Gürkan ÖZTÜRK.

ESE 523 Information Theory

Chapter 4 Multiple Random Variables

Lecture 3. Sampling, sampling distributions, and parameter estimation

Chapter 3 Sampling For Proportions and Percentages

Multivariate Transformation of Variables and Maximum Likelihood Estimation

Some Applications of the Resampling Methods in Computational Physics

Lecture 4 Sep 9, 2015

First Law of Thermodynamics

Generative classification models

Lecture 02: Bounding tail distributions of a random variable

Chain Rules for Entropy

BASICS ON DISTRIBUTIONS

CS 2750 Machine Learning Lecture 5. Density estimation. Density estimation

Econometrics. 3) Statistical properties of the OLS estimator

6. Nonparametric techniques

Econometric Methods. Review of Estimation

CODING & MODULATION Prof. Ing. Anton Čižmár, PhD.

X X X E[ ] E X E X. is the ()m n where the ( i,)th. j element is the mean of the ( i,)th., then

Parameter, Statistic and Random Samples

Mathematics HL and Further mathematics HL Formula booklet

Point Estimation: definition of estimators

Lecture Note to Rice Chapter 8

CS 1675 Introduction to Machine Learning Lecture 12 Support vector machines

Random Variables and Probability Distributions

Continuous Distributions

Introduction to local (nonparametric) density estimation. methods

9 U-STATISTICS. Eh =(m!) 1 Eh(X (1),..., X (m ) ) i.i.d

The expected value of a sum of random variables,, is the sum of the expected values:

Supplementary Material for Limits on Sparse Support Recovery via Linear Sketching with Random Expander Matrices

22 Nonparametric Methods.

Midterm Exam 1, section 1 (Solution) Thursday, February hour, 15 minutes

Lecture 7: Linear and quadratic classifiers

STK4011 and STK9011 Autumn 2016

Reliability evaluation of distribution network based on improved non. sequential Monte Carlo method

Introduction to Probability

best estimate (mean) for X uncertainty or error in the measurement (systematic, random or statistical) best

Binary classification: Support Vector Machines

Class 13,14 June 17, 19, 2015

Numerical Differentiation

Dimensionality Reduction and Learning

Random Variables. ECE 313 Probability with Engineering Applications Lecture 8 Professor Ravi K. Iyer University of Illinois

DISCRIMINATING BETWEEN WEIBULL AND LOG-LOGISTIC DISTRIBUTIONS

Lecture 9: Tolerant Testing

Bayesian belief networks

Analysis of System Performance IN2072 Chapter 5 Analysis of Non Markov Systems

Entropy, Relative Entropy and Mutual Information

Introduction to Matrices and Matrix Approach to Simple Linear Regression

Naïve Bayes MIT Course Notes Cynthia Rudin

STK3100 and STK4100 Autumn 2017

Overview. Basic concepts of Bayesian learning. Most probable model given data Coin tosses Linear regression Logistic regression

Section 2 Notes. Elizabeth Stone and Charles Wang. January 15, Expectation and Conditional Expectation of a Random Variable.

Bayes (Naïve or not) Classifiers: Generative Approach

INTERNATIONAL BACCALAUREATE ORGANIZATION GROUP 5 MATHEMATICS FORMULAE AND STATISTICAL TABLES

Normalized Exponential Tilting: Pricing and Measuring Multivariate Risks

Extreme Value Theory: An Introduction

7. Joint Distributions

Simple Linear Regression

Randomness and uncertainty play an important

D. VQ WITH 1ST-ORDER LOSSLESS CODING

Feature Selection: Part 2. 1 Greedy Algorithms (continued from the last lecture)

STA 105-M BASIC STATISTICS (This is a multiple choice paper.)

Multiple Linear Regression Analysis

Using Intraday Statistics for the Estimation of the Return Variance

1 Review and Overview

ECON 5360 Class Notes GMM

Transcription:

: Markku Jutt Overvew Te basc cocepts o ormato teory lke etropy mutual ormato ad EP are eeralzed or cotuous-valued radom varables by troduc deretal etropy ource Te materal s maly based o Capter 9 o te course book [] Telecomm Laboratory Course Overvew Basc cocepts ad tools Itroducto Etropy relatve etropy ad mutual ormato 3 symptotc equpartto t property 4 Etropy rates o a stocastc process ource cod or data compresso 5 Data compresso Cael capacty 8 Cael capacty 9 Deretal etropy 0 Te Gaussa cael Oter applcatos Mamum etropy ad spectral estmato 3 Rate dstorto teory 4 Network ormato teory Telecomm Laboratory Outle o te Lecture Revew o te last lecture Itroducto Detos Te EP or cotuous RV s Relato o deretal etropy ad dscrete etropy Jot ad codtoal deretal etropy Relatve etropy ad mutual ormato Propertes ummary Revew o Last Lecture Te udametal questos bed ormato teory: Wat s te ultmate data compresso rate? swer: etropy Wat s te ultmate data trasmsso rate? swer: cael capacty commucates wt B: pyscal acts o duce desred pyscal state B ad B sould aree o wat was set Te mamum umber o dstusable sals cael capacty = data rate lmt or relable commucato B Telecomm Laboratory 3 Telecomm Laboratory 4

Te Cael Cod Teorem ll rates below cael capacty C are acevable For all rate R < C: a sequece o ( R ) codes so tat 0 as (Weak) coverse: ay sequece o acevable ( R ) codes: R C tro coverse: or R < C P () e 0 ad R > C P () e epoetally Capacty wt eedback: supremum o acevable rates wt eedback codes C FB : C FB C ma I ; Y p Feedback does ot crease te capacty ource-cael Cod Teorem a source-cael code wt P () e 0 H(V)<C were te probablty o error s ˆ Pe PrV V pv p y v y v y v were Vˆ s te estmate o V s true (o - zero) () s a dcator ucto: 0 s alse (zero) Coverse: I H(V)>C probablty o error s bouded away rom zero ad relable trasmsso s ot possble Telecomm Laboratory 5 Telecomm Laboratory 6 ummary Itroducto Iormato cael capacty = operatoal cael capacty: C ma I; Y p ll rates below te capacty (R<C) are acevable: 0 as No rate above te capacty (R>C) s acevable Feedback does ot crease te capacty Feedback ca elp acev te capacty source wt etropy rate below te capacty ca be trasmtted relably source wt etropy rate above te capacty caot be trasmtted relably Dscrete-valued radom varables ave bee cosdered so ar Eteso or cotuous RV s eeded deretal etropy tratorward eteso o most oter cocepts partcular te asymptotc equpartto property ome mportat dereces tou Deretal etropy ca be eatve Dscrete etropy s always postve Mamum deretal etropy uder a covarace costrat by ormal (Gaussa) dstrbuto Dscrete etropy wt o costrat mamzed by uorm dstrbuto Telecomm Laboratory 7 Telecomm Laboratory 8

Detos Let be a RV wt cumulatve dstrbuto ucto (CDF) F() = Pr( ) I F() s cotuous te radom varable s sad to be cotuous Let df ' F I s ormalzed so tat d d () s called probablty desty ucto (PDF) or Te set o suc values o were () > 0 s called te support set (or smply support) o (or ucto ()) Deretal Etropy Deretal etropy ( te teral ests): d were s te support set o te radom varable () = E[()] Depeds o te PDF oly as does te etropy te dscrete case eeralzato o te etropy H p p marked also as () = () Telecomm Laboratory 9 Telecomm Laboratory 0 Eample #: Uorm Dstrbuto 0 a a /a 0 oterwse a a d a a a 0 d Note #: I a = (a) = 0 () = 0 Note #: I a < (a) < 0 Deretal etropy ca be eatve derece compared to te etropy However () = a = a s te volume o te support set wc always o-eatve Eample #: Normal Dstrbuto () Let te RV ollow ormal (e Gaussa) dstrbuto: ~ () () were deotes te PDF o real Gaussa dstrbuto wt mea zero ad varace : e Due to te epoet ucto t s smplest to calculate te deretal etropy ats Te deretal etropy bts ca be oud va cae o bases Telecomm Laboratory Telecomm Laboratory

Eample #: Normal Dstrbuto () l d l d E E l l l e l e ats e e l e bts e bts e e bts Te EP or Cotuous RV s Remd te asymptotc equpartto property or dscrete RV s: p H probablty Most o te probablty s cotaed te typcal set H : p : so tat p H p Pr or lare eou 3 H 3 Te same ca be sow or cotuous RV s Telecomm Laboratory 3 Telecomm Laboratory 4 EP ad Typcal et For a sequece o IID RV s: E probablty Proo: Follows drectly rom te weak law o lare umbers Typcal set: were Volume o set: Vol d d d R : Telecomm Laboratory 5 Propertes o te Typcal et Pr( () ) > or sucetly lare Vol( () ) (()+) or all 3 Vol( () ( ) (() ) ) or sucetly lare Proo: ee te tetbook [ ect 9] Te typcal set s te set wt te smallest volume satsy probablty bl ( ) to rst order o epoet Proo: mlar to te dscrete case Te smallest set tat cotas most o te probablty as appromately volume o Te correspod sde let s () = (sde let o m set wt most prob) Telecomm Laboratory 6

Relato o Deretal Etropy ad Dscrete Etropy Cosder a RV wt PDF () Let us dvde te rae o to bs o wdt ssume () s cotuous wt te bs By te mea value teorem: () : d Quatze: = <(+) Pr p d Telecomm Laboratory 7 Dervato Te etropy o te quatzed RV : H p p d d d I ()() ) s Rema terable d 0 Telecomm Laboratory 8 Teorem: Te Relato betwee te Deretal ad Dscrete Etropes I te PDF () s Rema terable te H 0 Te etropy o a -bt quatzato o a cotuous RV s appromately () + Eamples: I ~ U[0] ad = te =0adH( )= bts suce to descrbe to bt accuracy I ~ U[0/8] ad = te = 3 ad H( ) = Te rst 3 bts to te rt o decmal pot must be 0 3 bts suce to descrbe to bt accuracy I eeral () + s te averae umber o bts Te deretal etropy o a dscrete RV ca be cosdered to be result te volume o support set be zero Telecomm Laboratory 9 Jot ad Codtoal Deretal Etropy Jot deretal etropy: d d d Codtoal deretal etropy: Y y y d dy ( Y ) ( ) ( Y ) Telecomm Laboratory 0

Jot Deretal Etropy o a Multvarate Gaussa Radom Vector Te Gaussa PDF: ep T μ μ were = det() deotes te determat o te covarace matr Te deretal etropy o real-valued Gaussa RV s wt mea vector ad covarace matr : e Proo: ee te tetbook [ ect 94] Telecomm Laboratory Relatve Etropy ad Mutual Iormato Relatve etropy or ullback-lebler dstace: D Mutual ormato: y I ; Y y d dy y D y y Y Y Y Relato to dscrete mutual ormato: I ; Y H H Y Y I ;Y Telecomm Laboratory Propertes No-eatvty: D 0 D 0 almost everywere Proo by Jese s equalty: D 0 Furter propertes: I ; Y 0 I ; Y 0 y y Y Y y y Telecomm Laboratory 3 Propertes () Ca rule or deretal etropy: Idepedece boud: Hadamard equalty: For ~N(0): Traslatos ad rotatos: c a a Telecomm Laboratory 4

Te Gaussa PDF Mamzes Deretal Etropy Multvarate ormal (Gaussa) dstrbuto mamzes te deretal etropy over all dstrbutos wt te same covarace Let te real -dmesoal radom vector ave zero mea ad covarace = E(( T ) or j = E( ) j j Te e wt equalty ad oly ~N(0) Proo Let () be ay PDF satsy te covarace costrat jd j j Let be te PDF o N(0) or Now T ep ep 0 D * Telecomm Laboratory 5 Telecomm Laboratory 6 tep * te Proo T d e d T e T e d e e e d d ummary Eteso o ormato teoretc cocepts to cotuous RV s Deretal etropy: d symptotc equpartto property ad te typcal set: E probablty : Telecomm Laboratory 7 Telecomm Laboratory 8

ummary () Relatve etropy: D Mutual ormato: y I ; Y y d dy y D y y Y Y Y Multvarate ormal (Gaussa) dstrbuto mamzes te deretal etropy over all dstrbutos wt te same covarace: e Telecomm Laboratory 9