Concentration function and other stuff

Similar documents
Hamming Cube and Other Stuff

Randomized Algorithms

Outline. Martingales. Piotr Wojciechowski 1. 1 Lane Department of Computer Science and Electrical Engineering West Virginia University.

Lecture 5: Probabilistic tools and Applications II

Problem Set 1: Solutions Math 201A: Fall Problem 1. Let (X, d) be a metric space. (a) Prove the reverse triangle inequality: for every x, y, z X

Econ Lecture 3. Outline. 1. Metric Spaces and Normed Spaces 2. Convergence of Sequences in Metric Spaces 3. Sequences in R and R n

March 1, Florida State University. Concentration Inequalities: Martingale. Approach and Entropy Method. Lizhe Sun and Boning Yang.

On the Concentration of the Crest Factor for OFDM Signals

Course 212: Academic Year Section 1: Metric Spaces

Linear Analysis Lecture 5

An Algorithmist s Toolkit Nov. 10, Lecture 17

Lecture 1 Measure concentration

AN INEQUALITY FOR TAIL PROBABILITIES OF MARTINGALES WITH BOUNDED DIFFERENCES

Inequalities for Sums of Random Variables: a combinatorial perspective

Lecture I: Asymptotics for large GUE random matrices

Generalization Bounds and Stability

Polynomial Representations of Threshold Functions and Algorithmic Applications. Joint with Josh Alman (Stanford) and Timothy M.

A glimpse into convex geometry. A glimpse into convex geometry

Distance-Divergence Inequalities

Embeddings of finite metric spaces in Euclidean space: a probabilistic view


Definition 6.1. A metric space (X, d) is complete if every Cauchy sequence tends to a limit in X.

Generalization Bounds in Machine Learning. Presented by: Afshin Rostamizadeh

Entropy and Ergodic Theory Lecture 15: A first look at concentration

l 1 -Regularized Linear Regression: Persistence and Oracle Inequalities

Modelling self-organizing networks

Definable Extension Theorems in O-minimal Structures. Matthias Aschenbrenner University of California, Los Angeles

Problem Set 2: Solutions Math 201A: Fall 2016

A new approach to quantum metrics. Nik Weaver. (joint work with Greg Kuperberg, in progress)

THE SZEMERÉDI REGULARITY LEMMA AND ITS APPLICATION

Consistency of the maximum likelihood estimator for general hidden Markov models

Rademacher functions

SDS : Theoretical Statistics

1 Dimension Reduction in Euclidean Space

Introduction to Empirical Processes and Semiparametric Inference Lecture 12: Glivenko-Cantelli and Donsker Results

Mathematical Foundation for Compressed Sensing

MATH 31BH Homework 1 Solutions

STAT 200C: High-dimensional Statistics

Sliding Windows with Limited Storage

MATH 51H Section 4. October 16, Recall what it means for a function between metric spaces to be continuous:

Trace Reconstruction Revisited

Notions such as convergent sequence and Cauchy sequence make sense for any metric space. Convergent Sequences are Cauchy

16 Embeddings of the Euclidean metric

Reminder Notes for the Course on Measures on Topological Spaces

Lecture 3. Econ August 12

Class 2 & 3 Overfitting & Regularization

CMPSCI 711: More Advanced Algorithms

Math 4317 : Real Analysis I Mid-Term Exam 1 25 September 2012

Error Detection and Correction: Small Applications of Exclusive-Or

On the Uniqueness of Weak Solutions to the 2D Euler Equations

Based on the Appendix to B. Hasselblatt and A. Katok, A First Course in Dynamics, Cambridge University press,

A sequential hypothesis test based on a generalized Azuma inequality 1

Some Useful Background for Talk on the Fast Johnson-Lindenstrauss Transform

FUNCTIONAL ANALYSIS-NORMED SPACE

Introduction to Statistical Learning Theory

Math 320-2: Midterm 2 Practice Solutions Northwestern University, Winter 2015

AN EXPLORATION OF THE METRIZABILITY OF TOPOLOGICAL SPACES

The Moment Method; Convex Duality; and Large/Medium/Small Deviations

Measures. 1 Introduction. These preliminary lecture notes are partly based on textbooks by Athreya and Lahiri, Capinski and Kopp, and Folland.

THE SUBGAUSSIAN CONSTANT AND CONCENTRATION INEQUALITIES

Recent results in game theoretic mathematical finance

Talagrand's Inequality

functions as above. There is a unique non-empty compact set, i=1

Metric spaces and metrizability

Asymptotic Geometric Analysis, Fall 2006

Reminder Notes for the Course on Distribution Theory

Multichannel liar games with a fixed number of lies

Finite-dimensional spaces. C n is the space of n-tuples x = (x 1,..., x n ) of complex numbers. It is a Hilbert space with the inner product

2. Metric Spaces. 2.1 Definitions etc.

Uniform convergence of N-dimensional Walsh Fourier series

Lecture 3: AC 0, the switching lemma

Assignment 2 : Probabilistic Methods

Existence and Comparisons for BSDEs in general spaces

Lecture 10 Planck Distribution

Irredundant Families of Subcubes

Extended dynamic programming: technical details

Draft. Advanced Probability Theory (Fall 2017) J.P.Kim Dept. of Statistics. Finally modified at November 28, 2017

ABSTRACT INTEGRATION CHAPTER ONE

On the Concentration of the Crest Factor for OFDM Signals

Machine Learning Theory (CS 6783)

Chapter 9: Basic of Hypercontractivity

Some new estimates on the Liouville heat kernel

Empirical Risk Minimization

THE GREEN FUNCTION. Contents

B490 Mining the Big Data

Maths 212: Homework Solutions

Optimal compression of approximate Euclidean distances

Lecture 6 Proof for JL Lemma and Linear Dimensionality Reduction

OXPORD UNIVERSITY PRESS

Dynkin (λ-) and π-systems; monotone classes of sets, and of functions with some examples of application (mainly of a probabilistic flavor)

Densely defined non-closable curl on topologically one-dimensional Dirichlet metric measure spaces

Concentration Inequalities for Dependent Random Variables via the Martingale Method

STAT 200C: High-dimensional Statistics

An exponential separation between quantum and classical one-way communication complexity

The Completion of a Metric Space

Metric Spaces and Topology

Consistency of Modularity Clustering on Random Geometric Graphs

Lecture 18: March 15

DECOMPOSING DIFFEOMORPHISMS OF THE SPHERE. inf{d Y (f(x), f(y)) : d X (x, y) r} H

g 2 (x) (1/3)M 1 = (1/3)(2/3)M.

Transcription:

Concentration function and other stuff Sabrina Sixta Tuesday, June 16, 2014 Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 1 / 13

Table of Contents Outline 1 Chernoff Bound and Law of Large Numbers 2 Concentration Phenomenon 3 Blowing-up Lemma 4 Concentration function 5 Calculate value of concentration function for Hamming cube 6 Improvement on Chernoff Bounds 7 Hoeffding Inequality Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 2 / 13

Chernoff Bound and Law of Large Numbers Warming up to Chernoff Bound Let Ω = {A, B} be a two-element partition of a finite metric space X = (X, d) such that (1) d(a, B) = 1 and (2) there exists a bijection i : A B where d(a, i(a)) = 1 for all a A. Let f : X R be a 1-Lipschitz function. Then the function E[f Ω] : X R is 1-Lipschitz too. Or more relevent to us, E[f Ω](a) E[f Ω](b) 1. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 3 / 13

Chernoff Bound and Law of Large Numbers Warming up to Chernoff Bound Let Ω = {A, B} be a two-element partition of a finite metric space X = (X, d) such that (1) d(a, B) = 1 and (2) there exists a bijection i : A B where d(a, i(a)) = 1 for all a A. Let f : X R be a 1-Lipschitz function. Then the function E[f Ω] : X R is 1-Lipschitz too. Or more relevent to us, E[f Ω](a) E[f Ω](b) 1. Let f : {0, 1} n R be a 1-Lipschitz function relative to the Hamming distance, d and let f 1, f 2,..., f n be a corresponding martingale with respect to the standard refining partition. Then for every i 1, 2,... n, d i 1 2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 3 / 13

Chernoff Bound and Law of Large Numbers Chernoff Bound and Law of Large Numbers Chernoff Bound Let f : {0, 1} n R be a 1-Lipschitz function relative to the normalized Hamming distance, d. Then for every ɛ > 0, µ # {x {0, 1} n : f (x) E[f ] ɛ} 2e 2nɛ2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 4 / 13

Chernoff Bound and Law of Large Numbers Chernoff Bound and Law of Large Numbers Chernoff Bound Let f : {0, 1} n R be a 1-Lipschitz function relative to the normalized Hamming distance, d. Then for every ɛ > 0, µ # {x {0, 1} n : f (x) E[f ] ɛ} 2e 2nɛ2. Law of Large Numbers Applying the normalized weight function, w, to the Chernoff bound, we get for ɛ > 0, µ # {x {0, 1} n : w 1 2 ɛ} 2e 2nɛ2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 4 / 13

Chernoff Bound and Law of Large Numbers Chernoff Bound and Law of Large Numbers cont. Figure: Actual concentration values for the Hamming Cube, {0, 1} 10 vs. the Chernoff Bound. µ # {x {0, 1} n : w 1 2 n 2k 2n } = 2 2 n k i=0 ( ) n. i Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 5 / 13

Concentration Phenomenon Concentration Phenomenon Azuma s inequality and the Chernoff bounds are both part of a larger group of inequalities that manifest the phenomenon of concentration of measure on structures of high dimension. That is, on a typical mathematical structure of high dimension (like the Hamming cube, {0, 1} n ) every nice function (1-Lipschitz functions) concentrates near one value. Bounds like Chernoff s are called concentration inequalities. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 6 / 13

Concentration Phenomenon Concentration Phenomenon Cont. Figure: Orthogonal projection of 1000 random points on a unit Euclidean d-cube (clockwise, d=3,10,100,1000). Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 7 / 13

Blowing-up Lemma Warming up to Blowing-up Lemma ɛ-neighbourhood Let A X, where (X, d) is a metric space and let ɛ > 0. The ɛ-neighbourhood of A in X is, A ɛ = {x X : there exists a A, d(x, a) < ɛ}. Extended Chernoff inequality Let f : {0, 1} n R be a 1-Lipschitz function relative to the normalized Hamming distance. Then for every ɛ > 0 µ # {x {0, 1} n : f (x) E[f ] ɛ} e 2nɛ2 and µ # {x {0, 1} n : f (x) E[f ] ɛ} e 2nɛ2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 8 / 13

Blowing-up Lemma Blowing-up Lemma Blowing-up lemma Let A {0, 1} n be a set of n-bit strings such that µ # (A) 1 2, then for ɛ > 0, the ɛ-neighbourhood for A, with respect to the normalized Hamming distance, satisfies the following inequality, µ # (A ɛ ) 1 e ɛ2 n/2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 9 / 13

Concentration function Concentration function The concentration function of the Hamming cube {0, 1} n, denoted α(n, ɛ) or α({0, 1} n, ɛ) on R +, is { 1 α(ɛ) = 2 if ɛ = 0 1 min{µ # (B ɛ ) : B {0, 1} n, µ # (B) 1 2 } if ɛ > 0. (1) Some properties of the concentration function on the Hamming cube, {0, 1} n : non-increasing, left-continuous, not right-continuous. The concentration function of the Hamming cube {0, 1} n satisfies α({0, 1} n, ɛ) e ɛ2 n/2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 10 / 13

Calculate value of concentration function for Hamming cube Calculate value of concentration function for Hamming cube Hamming ball A set B {0, 1} n is called a Hamming ball if for σ {0, 1} n and k = 1, 2,..., n, B k (σ) B B k+1 (σ). Harper isopetrimetric theorem For every A in the metric space ({0, 1} n, d) and each k = 1, 2,..., n there is a Hamming ball B with #(B) = #(A) and #(B k ) #(A k ). For a Hamming cube of odd dimension, {0, 1} 2m+1 equipped with the normalized distance d, the value of the concentration function for ɛ > 0 is where n = 2m + 1. m+nɛ α(ɛ) = 1 2 n k=0 ( n k ) n m nɛ = 2 n k=0 ( ) n k Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 11 / 13

Improvement on Chernoff Bounds Improvement on Chernoff Bounds The concentration function of the Hamming cube, {0, 1} n, satisfies the following inequality α({0, 1} n, ɛ) e 2ne2. Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 12 / 13

Hoeffding Inequality Hoeffding Inequality Rademach s Cube: { 1, 1} n = {σ 1, σ 2,..., σ n : σ i { 1, 1} for i 1, 0,..., n} Hoeffding Inequality This inequality is another modification of Azuma s inequality for the Rademach s Cube. Let a = (a 1, a 2,..., a n ) R n then for c > 0, µ # {η { 1, 1} n : n i=1 ( η i a i > c} < 2exp c 2 a 2 2 ). Sabrina Sixta () Concentration function and other stuff Tuesday, June 16, 2014 13 / 13