Lecture 4 Sep 9, 2015

Similar documents
The Occupancy and Coupon Collector problems

Lecture 02: Bounding tail distributions of a random variable

Law of Large Numbers

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

Homework 1: Solutions Sid Banerjee Problem 1: (Practice with Asymptotic Notation) ORIE 4520: Stochastics at Scale Fall 2015

,m = 1,...,n; 2 ; p m (1 p) n m,m = 0,...,n; E[X] = np; n! e λ,n 0; E[X] = λ.

Introduction to Probability

X ε ) = 0, or equivalently, lim

Complete Convergence and Some Maximal Inequalities for Weighted Sums of Random Variables

D KL (P Q) := p i ln p i q i

Econometric Methods. Review of Estimation

Chapter 5 Properties of a Random Sample

Class 13,14 June 17, 19, 2015

18.657: Mathematics of Machine Learning

3. Basic Concepts: Consequences and Properties

Lecture 3 Probability review (cont d)

Lecture 9: Tolerant Testing

hp calculators HP 30S Statistics Averages and Standard Deviations Average and Standard Deviation Practice Finding Averages and Standard Deviations

ρ < 1 be five real numbers. The

UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS

Chapter 4 Multiple Random Variables

CHAPTER VI Statistical Analysis of Experimental Data

1 Solution to Problem 6.40

PTAS for Bin-Packing

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

Dimensionality Reduction and Learning

Point Estimation: definition of estimators

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

CIS 800/002 The Algorithmic Foundations of Data Privacy October 13, Lecture 9. Database Update Algorithms: Multiplicative Weights

CS286.2 Lecture 4: Dinur s Proof of the PCP Theorem

Parameter, Statistic and Random Samples

Bayes (Naïve or not) Classifiers: Generative Approach

UNIVERSITY OF OSLO DEPARTMENT OF ECONOMICS

Runtime analysis RLS on OneMax. Heuristic Optimization

Third handout: On the Gini Index

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

Functions of Random Variables

8.1 Hashing Algorithms

Extend the Borel-Cantelli Lemma to Sequences of. Non-Independent Random Variables

Chapter 5 Properties of a Random Sample

Summary of the lecture in Biostatistics

The Arithmetic-Geometric mean inequality in an external formula. Yuki Seo. October 23, 2012

Chain Rules for Entropy

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

Bounds on the expected entropy and KL-divergence of sampled multinomial distributions. Brandon C. Roy

Special Instructions / Useful Data

THE ROYAL STATISTICAL SOCIETY 2016 EXAMINATIONS SOLUTIONS HIGHER CERTIFICATE MODULE 5

Chapter 8: Statistical Analysis of Simulated Data

Chapter 4 Multiple Random Variables

THE ROYAL STATISTICAL SOCIETY GRADUATE DIPLOMA

A tighter lower bound on the circuit size of the hardest Boolean functions

Rademacher Complexity. Examples

Hard Core Predicates: How to encrypt? Recap

Lecture 2 - What are component and system reliability and how it can be improved?

Extreme Value Theory: An Introduction

Arithmetic Mean Suppose there is only a finite number N of items in the system of interest. Then the population arithmetic mean is

Multiple Choice Test. Chapter Adequacy of Models for Regression

Estimation of Stress- Strength Reliability model using finite mixture of exponential distributions

Complete Convergence for Weighted Sums of Arrays of Rowwise Asymptotically Almost Negative Associated Random Variables

Module 7. Lecture 7: Statistical parameter estimation

Continuous Distributions

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

Chapter 14 Logistic Regression Models

Lecture Notes Types of economic variables

Lecture Note to Rice Chapter 8

Pr[X (p + t)n] e D KL(p+t p)n.

STK4011 and STK9011 Autumn 2016

Algorithms behind the Correlation Setting Window

PROJECTION PROBLEM FOR REGULAR POLYGONS

Q-analogue of a Linear Transformation Preserving Log-concavity

Introduction to Matrices and Matrix Approach to Simple Linear Regression

Exchangeable Sequences, Laws of Large Numbers, and the Mortgage Crisis.

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

1 Review and Overview

å 1 13 Practice Final Examination Solutions - = CS109 Dec 5, 2018

Non-uniform Turán-type problems

A Study of the Reproducibility of Measurements with HUR Leg Extension/Curl Research Line

Arithmetic Mean and Geometric Mean

Application of Generating Functions to the Theory of Success Runs

Qualifying Exam Statistical Theory Problem Solutions August 2005

Introduction to local (nonparametric) density estimation. methods

Algorithms Design & Analysis. Hash Tables

Arora Rao Vazirani Approximation for Expansion

Likewise, properties of the optimal policy for equipment replacement & maintenance problems can be used to reduce the computation.

Applying the condition for equilibrium to this equilibrium, we get (1) n i i =, r G and 5 i

NP!= P. By Liu Ran. Table of Contents. The P versus NP problem is a major unsolved problem in computer

5 Short Proofs of Simplified Stirling s Approximation

Ordinary Least Squares Regression. Simple Regression. Algebra and Assumptions.

1. The weight of six Golden Retrievers is 66, 61, 70, 67, 92 and 66 pounds. The weight of six Labrador Retrievers is 54, 60, 72, 78, 84 and 67.

Mean is only appropriate for interval or ratio scales, not ordinal or nominal.

Lecture 9. Some Useful Discrete Distributions. Some Useful Discrete Distributions. The observations generated by different experiments have

NP!= P. By Liu Ran. Table of Contents. The P vs. NP problem is a major unsolved problem in computer

Lecture 3. Sampling, sampling distributions, and parameter estimation

Marcinkiewicz strong laws for linear statistics of ρ -mixing sequences of random variables

Beam Warming Second-Order Upwind Method

{ }{ ( )} (, ) = ( ) ( ) ( ) Chapter 14 Exercises in Sampling Theory. Exercise 1 (Simple random sampling): Solution:

Mu Sequences/Series Solutions National Convention 2014

PART ONE. Solutions to Exercises

Investigation of Partially Conditional RP Model with Response Error. Ed Stanek

MEASURES OF DISPERSION

Transcription:

CS 388R: Radomzed Algorthms Fall 205 Prof. Erc Prce Lecture 4 Sep 9, 205 Scrbe: Xagru Huag & Chad Voegele Overvew I prevous lectures, we troduced some basc probablty, the Cheroff boud, the coupo collector problem, ad game tree evaluato. I ths lecture, we wll troduce cocetrato equaltes. 2 Coupo Collector Problem Draw umbers (coupos) depedetly from ] = {, 2,..., }. How log does t take to see all of the umbers? Suppose T s the umber of draws to get the -th ew umber. Let T = T. Fact. The T s are depedet of each other. Fact 2. T follows geometrc dstrbuto wth success probablty, p = +. Fact 3. If X Geometrc(p), EX] = p + ( p) (EX X 2]) EX] = p = p + ( p)( + EX]) EX 2 ] = p 2 + ( p) EX 2 X 2] = p + ( p)e(x + ) 2 ] = p + ( p)(ex 2 ] + 2EX] + ) = p + ( p)ex 2 ] + 2( p)/p + ( p) EX 2 ] = 2 p p 2 V ar(x) = E(X E(X)) 2 ] = EX 2 ] (EX]) 2 V ar(x) = p p 2 p 2

Therefore the Coupo Collector Problem, ET ] = ET ] = = = + = H log V art ] ( p 2 = + V art ] = ) 2 = V art ] 2 ( = ) 2 2 π2 6 = O(2 ) 3 Cocetrato Iequaltes, P rt + α] Assume s some falure probablty. Settg α = have Defto 4. Uo Boud ( + ( +, P rt + α ] ) log ) α ad because ( x )x < e, we P rx X 2... X ] P rx ] Usg a uo boud, we have P r T + log log ] = P r T + ] α = P rt + α... T + α ] P rt + α ] Defto 5. Wth Hgh Probablty (w.h.p.) X O(y) w.h.p. c 2, c, s.t. P rx c y] c 2 T = O( log 2 ) wth hgh probablty. 2

3. Markov s Iequalty For a o-egatve radom varable T ad ay o-egatve α, I the Coupo Collector Problem, ET ] P rt α] α P rt α] ET ] α α = ET ] P r = H T H ] 3.2 Chebyshev s Iequalty For a radom varable, X, let µ = EX] deote the expectato ad σ 2 = V arx] deote the varace. Startg from Markov s Iequalty, we fd Settg α ασ P r(x µ) 2 α 2 ] E(X µ)]2 α 2 = σ2 α 2 P r(x µ) 2 α 2 σ 2 ] α 2 Takg the square root, we fd P rx µ + ασ] α 2 P rx µ ασ] α 2 Usg ths result the Coupo Collector Problem, gves us P rt H + O()] Settg = log 2 Most of the tme, the typcal devato s O(σ). ( ) P rt H + O( log )] O log 2 P r x µ O(σ)] 3

3.3 Momet Method If f s o-egatve, by Markov s equalty, P rf(x µ) f(α)] Ef(X µ)] f(α) For f creasg, Set f = t k, For oe sde, P rx µ α] Ef(X µ)] f(α) P r X µ k α k ] E x µ k ] α k P rx µ + α] E x µ k ] α k Settg = E x µ k ], we have α k P r X µ + E x µ k ] /k ( ) ] /k If we cosder X N(0, σ 2 ), we kow E x k ] (kσ 2 ) k/2 k > 0 whch meas Settg k = log, we get P r X µ + O P r X µ + O ( ( ) )] k /k σ ( log )] 3.4 Momet Geeratg Fucto Defto 6. The momet geeratg fucto, parameterzed by λ, s defed as MGF X (λ) = Ee λ(x µ) ] Assume X s cetered (EX] = 0). e λx = + λx + (λx)2 2 + (λx)3 3! +... + (λx)k k! We ca use parameter λ to adjust the weghts o each term. Whe λ s larger, more weght s o hgher order terms. 4

From the dervato of the Momet Method, settg f(x) = e λx, P rx µ + α] MGF X(λ) e λα Fact 7. If X N(0, σ 2 ), MGF X (λ) = Ee λx ] e λ2 σ 2 Usg ths, Set λ = α σ 2, we get If = e α2 2σ 2, we have λ R P rx µ + α] MGF x(λ) e λα = e λ2 σ 2 2 λα = e 2(λσ α σ ) 2 α2 2σ 2 P rx µ + α] e α2 2σ 2 α = σ 2log ( ) Note that ths s the same O log boud as we foud the method of momets, except that ow we kow the costat. 3.5 Subgaussa Varables Clam 8. The followg three statemets are equvalet f we oly care up to a costat for σ (.e., j {, 2, 3}, σ = θ(σ j )) X s subgaussa wth parameter σ,.e. λ R, MGF X (λ) e λ 2 σ 2 2 () P rx µ + t] e t 2 ) E x k ] /k O (σ 3 k Fact 9. The sum of subgaussa varables are subgaussa. 2σ 2 2 (2) X = X +... + X MGF X (λ) = E e λx] ] = E e λ( x ) = E = ] E e λx (by depedece) e λx ] (3) = MGF X (λ) e λ2 σ 2 /2 (by subgaussa) = e λ2 2 ( σ2 ) 5

Ths mples X s subgaussa wth parameter σ2. Fact 0. If X 0, ], the X s subgaussa wth σ = /2 by Hoeffdg s Lemma. Let X = X where X 0, ]. X s subgaussa wth σ = /2. Plug ths to (2), we have whch s exactly the Cheroff boud. P rx µ + α] e 2α2 4 Next Class I the Coupo Collector Problem, we had P rt α] ( ) α e α Ths s ot of the form e α2 so we caot use the subgaussa results. We wll relax the subgaussa requremet to subexpoetal ad subgamma. Ths wll lead to Berste s equalty. MGF X (λ) = Ee λx ] e λ2 σ 2 2 λ B Refereces MR] Rajeev Motwa, Prabhakar Raghava Radomzed Algorthms. Cambrdge Uversty Press, 0-52-47465-5, 995. RV] Roma Vershy Itroducto to the o-asymptotc aalyss of radom matrces. CoRR, abs-0-3027, 200. https://e.wkpeda.org/wk/hoeffdg%27s_equalty 6