Problem Set 2 Solutions

Similar documents
Randomized Algorithms I, Spring 2018, Department of Computer Science, University of Helsinki Homework 1: Solutions (Discussed January 25, 2018)

n outcome is (+1,+1, 1,..., 1). Let the r.v. X denote our position (relative to our starting point 0) after n moves. Thus X = X 1 + X 2 + +X n,

Lecture 2. The Lovász Local Lemma

Discrete Mathematics for CS Spring 2007 Luca Trevisan Lecture 22

HOMEWORK 2 SOLUTIONS

Lecture 2 February 8, 2016

Math 216A Notes, Week 5

Lecture 2: April 3, 2013

An Introduction to Randomized Algorithms

Lecture 4: April 10, 2013

Basics of Probability Theory (for Theory of Computation courses)

Lecture 2: Concentration Bounds

Discrete Mathematics for CS Spring 2005 Clancy/Wagner Notes 21. Some Important Distributions

Math 155 (Lecture 3)

ACO Comprehensive Exam 9 October 2007 Student code A. 1. Graph Theory

Statistical and Mathematical Methods DS-GA 1002 December 8, Sample Final Problems Solutions

Ma/CS 6b Class 19: Extremal Graph Theory

On Random Line Segments in the Unit Square

Discrete Mathematics and Probability Theory Summer 2014 James Cook Note 15

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 19 11/17/2008 LAWS OF LARGE NUMBERS II THE STRONG LAW OF LARGE NUMBERS

Application to Random Graphs

Lecture 01: the Central Limit Theorem. 1 Central Limit Theorem for i.i.d. random variables

1 Last time: similar and diagonalizable matrices

7.1 Convergence of sequences of random variables

Learning Theory: Lecture Notes

Lecture 12: November 13, 2018

Discrete Mathematics and Probability Theory Spring 2016 Rao and Walrand Note 19

Lecture 4: Unique-SAT, Parity-SAT, and Approximate Counting

Lecture 3: August 31

7.1 Convergence of sequences of random variables

Discrete Mathematics for CS Spring 2008 David Wagner Note 22

Notes for Lecture 11

CS / MCS 401 Homework 3 grader solutions

Discrete Mathematics and Probability Theory Spring 2012 Alistair Sinclair Note 15

Math 61CM - Solutions to homework 3

Seunghee Ye Ma 8: Week 5 Oct 28

Linear regression. Daniel Hsu (COMS 4771) (y i x T i β)2 2πσ. 2 2σ 2. 1 n. (x T i β y i ) 2. 1 ˆβ arg min. β R n d

Lecture 14: Graph Entropy

THE ASYMPTOTIC COMPLEXITY OF MATRIX REDUCTION OVER FINITE FIELDS

Design and Analysis of Algorithms

Disjoint Systems. Abstract

The multiplicative structure of finite field and a construction of LRC

Bertrand s Postulate

ECE 901 Lecture 12: Complexity Regularization and the Squared Loss

The picture in figure 1.1 helps us to see that the area represents the distance traveled. Figure 1: Area represents distance travelled

Discrete Mathematics and Probability Theory Fall 2016 Seshia and Walrand Final Solutions

STAT Homework 1 - Solutions

CS 171 Lecture Outline October 09, 2008

Large holes in quasi-random graphs

Lecture 5: April 17, 2013

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 2 9/9/2013. Large Deviations for i.i.d. Random Variables

Math 778S Spectral Graph Theory Handout #3: Eigenvalues of Adjacency Matrix

Chapter 2 The Solution of Numerical Algebraic and Transcendental Equations

Eigenvalues and Eigenvectors

Ada Boost, Risk Bounds, Concentration Inequalities. 1 AdaBoost and Estimates of Conditional Probabilities

Supplemental Material: Proofs

w (1) ˆx w (1) x (1) /ρ and w (2) ˆx w (2) x (2) /ρ.

f X (12) = Pr(X = 12) = Pr({(6, 6)}) = 1/36

Advanced Stochastic Processes.

Random Matrices with Blocks of Intermediate Scale Strongly Correlated Band Matrices

Convergence of random variables. (telegram style notes) P.J.C. Spreij

Notes 19 : Martingale CLT

Lecture 6: Coupon Collector s problem

Lecture 9: Expanders Part 2, Extractors

Complex Numbers Solutions

Topic 9: Sampling Distributions of Estimators

Problem Set 4 Due Oct, 12

Fall 2013 MTH431/531 Real analysis Section Notes

1 Generating functions for balls in boxes

Solutions to Final Exam

Random Models. Tusheng Zhang. February 14, 2013

Mathematical Foundations -1- Sets and Sequences. Sets and Sequences

Solutions to selected exercise of Randomized Algorithms

Lecture 10 October Minimaxity and least favorable prior sequences

Problems from 9th edition of Probability and Statistical Inference by Hogg, Tanis and Zimmerman:

1 Hash tables. 1.1 Implementation

Discrete Mathematics and Probability Theory Spring 2014 Anant Sahai Lecture 16

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 3 9/11/2013. Large deviations Theory. Cramér s Theorem

Some remarks for codes and lattices over imaginary quadratic

Machine Learning Brett Bernstein

The Random Walk For Dummies

Shannon s noiseless coding theorem

Topic 9: Sampling Distributions of Estimators

CHAPTER 5. Theory and Solution Using Matrix Techniques

Probabilistic Methods in Combinatorics

Math 220B Final Exam Solutions March 18, 2002

Math 113, Calculus II Winter 2007 Final Exam Solutions

MATH 304: MIDTERM EXAM SOLUTIONS

10-701/ Machine Learning Mid-term Exam Solution

Chapter 5. Inequalities. 5.1 The Markov and Chebyshev inequalities

This exam contains 19 pages (including this cover page) and 10 questions. A Formulae sheet is provided with the exam.

Math 4707 Spring 2018 (Darij Grinberg): homework set 4 page 1

Generalized Semi- Markov Processes (GSMP)

Lecture 12: September 27

Singular Continuous Measures by Michael Pejic 5/14/10

6.883: Online Methods in Machine Learning Alexander Rakhlin

Stochastic Matrices in a Finite Field

CS 330 Discussion - Probability

6.895 Essential Coding Theory October 20, Lecture 11. This lecture is focused in comparisons of the following properties/parameters of a code:

Notes 5 : More on the a.s. convergence of sums

Transcription:

CS271 Radomess & Computatio, Sprig 2018 Problem Set 2 Solutios Poit totals are i the margi; the maximum total umber of poits was 52. 1. Probabilistic method for domiatig sets 6pts Pick a radom subset S of vertices by icludig each vertex of V i S idepedetly with probability p (a value to be chose later). Let T U be the set of vertices i U that are either i S or adjacet to a vertex of S. The clearly S T is a domiatig set for U. Note that a vertex u U belogs to T iff either it or ay of its (at least d) eighbors belogs to S, which happes with probability at most (1 p). Hece the expected size of the domiatig set is E( S T ) = E( S ) + E( T ) p + (1 p) (p + e p() ). Now we choose p so as to miimize this expressio. Differetiatig the term i the paretheses, we see that we wat to take e p() = 1 l(), or p =. Pluggig this ito the above expressio gives E( S T ) log()+1, so we deduce that there must exist a domiatig set of at most this size. Ideed, sice the size of a domiatig set must be a iteger, we kow there exists a domiatig set of size log()+1. 2. Locally 2-colorable graphs (a) Let the r.v. X deote the umber of edges i G. Sice X is the sum of ( idicator r.v. s Xe (oe for each possible edge e), we have by liearity of expectatio EX = ( p = 8( 1). Moreover, sice the X e are idepedet, Var(X) = e Var(X e) = ( p(1 p) < EX. Now Chebyshev s iequality gives Pr[X > 10( 1)] Pr [ X EX > 1 4 EX] Var(X) (EX/4) 2 = 16 EX = 2 1. This fractio is less tha 1 4 for all 9. (b) Fix a colorig; suppose it has r red vertices ad r gree vertices. Let the r.v. Y deote the umber of violated edges. The by liearity EY = ( ) ( r 2 + r ) 2 )p. This expressio is miimized whe r = /2 (assumig is eve), givig EY 2 ( ) /2 2 p = 4(. Now applyig the Cheroff boud to Y with δ = 4 gives Pr[Y 1] Pr[Y 1 4 EY ] exp( 1 2 ( 4 )2 4( ) = exp ( 9 8 ( ). (c) Call a 2-colorig good if it has at most 2 violated edges. Part (b) bouds the probability that ay give colorig is good. Takig a uio boud over all 2 possible colorigs, we get where the last iequality holds for all 9. Pr[ a good colorig for G] 2 exp ( 9 8 ( ) 1 4, Commo Issues: May people forgot to take the uio boud over the 2 possible colorigs. Part (b) oly applies to oe fixed colorig, so this is ecessary.

(d) Let the r.v. Z deote the umber of k-cycles i G. The we ca write Z = C Z C, where the sum is over all possible k-cycles C ad Z C is the idicator r.v. for the evet that C is a cycle i G. Sice C cosists of k edges, we have Pr[Z C = 1] = p k. Also, the umber of possible C is ( ) k! ) k 2k ; here is the umber of ways of choosig the vertices i C, while k! is the umber of vertex orderigs ( k ad 2k accouts for the choice of startig poit ad the sese of traversal of the cycle. By liearity of expectatio, we have EZ = C EZ C = ( ) k! k 2k pk (p) k = 16 k. The expected umber of cycles of legth at most 1 8 log is thus bouded by 1 8 log k= 16 k 16 15 16 1 8 log = 16 15 < 2. (e) By part (d) ad Markov s iequality, we have that Pr[Z 8 ] 1 4. So with probability at least 4, G has at most 8 cycles of legth at most 1 8 log. Now if we delete a (arbitrary) edge from each of these cycles (a total of at most 8 edges), we are left with a graph whose shortest cycle is loger tha 1 8 log. But this implies that ay iduced subgraph o up to 1 8 log vertices is cycle-free, ad hece certaily 2-colorable. (f) Takig a uio boud over the evets i parts (a), (c) ad (e), we deduce that, with positive probability (at least 1 4 ), the radom graph G has all of those properties simultaeously. For such a G, let G be the graph obtaied by deletig at most 8 edges from G, as i part (e). The G has all the followig properties: (i) G has at most 10( 1) edges (by part (a)); (ii) G is ot 2-colorable eve if we delete ay 2 8 edges (by part (c); ote that whe deletig edges to remove cycles we may remove some violated edges); (iii) the iduced subgraph o ay 1 8 log vertices of G is 2-colorable (by part (e)). Fially, observe that (very crudely) 2 8 > 1 1 2 ( 1), which by (i) is at least 20 of the umber of edges i G. Commo Issues: May people failed to take a uio boud over parts (a), (c) ad (e), istead claimig these evets were idepedet. May people also forgot to accout for the 8 edges that were deleted via part (c).. A threshold for isolated vertices (a) For i = 1,...,, let X i be the idicator radom variable for the evet that vertex i is isolated, ad let X = i X i be the umber of isolated vertices. To show that p = l is a threshold, we eed to show the followig two facts: 1. if p l, the Pr[X > 0] 0 as. 2. if p l, the Pr[X > 0] 1 as. [Here, as i lectures, for fuctios f(), g(), f g meas that f()/g() as, ad f g meas that f()/g() 0 as.] Clearly EX = (1 p) 1 =: µ. Note that 8pts l µ = l + ( 1) l(1 p) l ( 1)p. (1) [Here f g meas that f()/g() 1 as.] Thus, as, we have µ 0 if p l ad µ if p l For p l. we immediately have by Markov s iequality that which establishes Fact 1 above. Pr[X > 0] = Pr[X 1] µ 0 as,,

To establish Fact 2, for p l we eed to use the secod momet method, as follows: Pr[X = 0] Pr[ X µ µ] VarX µ 2. Thus it is sufficiet to show that VarX µ 2 0. To do this, ote that EX 2 = i EX 2 i + i j E[X i X j ] = µ + ( 1)E[X i X j ]. Also, for ay i j we have E[X i X j ] = (1 p) 2, ad thus ( 1)E[X i X j ] µ 2 /(1 p). Hece sice µ ad p 0. This establishes Fact 2. VarX µ 2 µ + µ2 /(1 p) µ 2 µ 2 = 1 µ + p 0, ( 1 p (b) If we set p = c l the the above calculatios carry over essetially uchaged. I particular, from (1) above we see that µ 0 if c > 1 ad µ if c < 1. Thus by Markov s iequality we still get that Pr[X > 0] 0 i the former case. Ad i the latter case the calculatio i ( still holds as well, so we get Pr[X = 0] 0 i this case. This shows that the threshold is sharp; the Note i the problem set tells you that the width of the scalig widow is actually 1. 4. Plated cliques ad cryptography (a) Cosider the set of pairs (C, G) where C is a set of k vertices, ad G is a graph. Each pair uiquely correspods to a specific derivatio of a graph i G,1/2, where G is the graph chose accordig to G,1/2, ad C is the set of k vertices chose to become a clique. Clearly each pair is equally likely, arisig with probability ( k) 12 (, where the first term is the probability of selectig C, ad the secod is the probability of choosig G, sice all graphs are equally likely i G,1/2. For a give graph G, the umber of pairs (C, G) which would give rise to it is f(g )2 (k, sice C must be oe of the cliques of G, ad give the choice of C, G ca have ay subset of the ( k edges betwee vertices i C. Thus Pr [G ] = f(g )2 (k ( ) 12 ( k = f(g ) Pr[G ] µ. Commo Issues: May people simply claimed that the umber of graphs G that could have give rise to a graph G (i G,1/ is f(g )2 (k. This is ot true: to see a silly example, let G be a graph cosistig of two disjoit k-cliques; the the umber of graphs that could have give rise to G is 2 2 (k 1. To see why the 1 is there, ote that for either choice of the clique to add, there are 2 ( k possible graphs that could have yielded G ; however, we are double-coutig the graph G, sice that graph could have give rise to G uder either choice of which clique to add (because o ew edges are eeded). Whe coutig, oe must be careful to avoid double-coutig! (b) By Chebyshev s iequality, we have Pr[G is α-bad] Pr[ f(g) µ > ( α 1)µ] E(f 2 ) µ 2 ( α 1) 2 = O(c log ) ( α 1) 2, where i the last step we used the give upper boud o E(f 2 ) µ 2. For large eough, this last expressio is bouded above by 2α+c+ɛ for ay ɛ > 0. (c) For B j defied as i the hit, we have Pr [G is α-bad] = Pr [G B j ]

(j+1)α/2 Pr[G B j ] (j+1)α/2 Pr[f(G) > jα/2 ] (j+1)α/2 jα+c+ɛ = α/2+c+ɛ jα/2 = α/2+c+ɛ O( α ) = O( α/2+c+ɛ ). The secod lie here follows from part (a), ad the fourth lie from part (b). (d) We assume that Pr [E holds for G] = Ω( r ). Now ote that Pr[E holds for G] Pr[E holds for G ad G is ot α-bad] α Pr [E holds for G ad G is ot α-bad] α ( Pr [E holds for G] Pr [G is α-bad] ) ( ) = α Ω( r ) O( α/2+c+ɛ ). (The secod lie here uses part (a), ad the last lie part (c).) Fially, we are free to choose α so that α/2 c ɛ > r, which implies that the egative term i the above fial expressio is o( r ). We coclude that Pr[E holds for G] = Ω( r α ) = Ω( r ), as desired. Commo Issues: May people failed to take a uio boud over the two evets (E holds i G) ad (G is ot α-bad), istead claimig icorrectly that these evets were idepedet. 5. More o ubalacig lights (a) Sice z 0 the required iequality, for ay fixed α > 0, is equivalet to z z α 2 α. Differetiatig the left-had side ad settig to zero, we get a turig poit at z = α, ad sice the secod derivative is egative this must be a maximum. There are o other turig poits i z 0, so the global maximum value is 2 α, as desired. For ay radom variable Z, applyig the above iequality to Z ad takig expectatios we get E( Z ) ( ) 2 E(Z 2 ) E(Z4 ). α α This holds for ay α > 0, so we may choose α to maximize the right-had side. (We may also assume that E(Z 2 ) 0; otherwise Z = 0 with probability 1.) Differetiatig, we see that the optimal α is α = E(Z4 ). Pluggig this i ad tidyig up gives the claimed boud. E(Z 2 ) (b) Settig Z = S i part (a), we eed to evaluate E(S 2 ) ad E(S 4 ). First we have E(S 2 ) = i E(X 2 i ) + i j E(X i X j ) =.

To see this, ote that E(Xi = 1 sice X i is ±1-valued; ad E(X i X j ) = E(X i )E(X j ) = 0 sice the X i are 4-wise idepedet (ad hece certaily pairwise idepedet). By similar cosideratios we have E(S) 4 = E(Xi 4 ) + 1 ( ) 4 E(Xi 2 Xj 2 ) = + ( 1) = 2 2. 2 2 i i j Note that all other terms have expectatio zero because of 4-wise idepedece ad the fact that E(X i ) = 0. (The 1 4 2( = arises as the umber of ways of partitioig the two copies of Xi ad X j amog the four factors.) Pluggig these two values ito the boud from part (a) gives /2 E( S ) ( 2 1/2. (c) First, suppose we use the radomized algorithm of Lecture 5, except that ow the coi tosses used to determie the colum switch settigs are oly 4-wise idepedet. The aalysis i Lecture 5 relied 2 oly o the fact that E( Z i ) is asymptotically at least π, where Z i = j=1 X ij is the sum of light values for row i. Sice the X ij are also 4-wise idepedet, we may substitute the result of part (b) to deduce that E( Z i ) is asymptotically at least, ad hece the asymptotic expected excess is 1 /2 1. (Note that 2 0.58 while π 0.80.) Now recall from Lecture 10 that we ca costruct a family of r d-wise idepedet fair coi flips usig radom polyomials of degree d 1 over a field of size q = O(r). (Say q = 2 m, where m = log 2 r.) I our preset applicatio, r = ad d = 4. The umber of poits i this sample space (i.e., the umber of such polyomials) is oly q d = O( 4 ), so istead of ruig the radomized algorithm we may exhaustively try all sample poits i polyomial time. We are guarateed to fid at least oe sample poit for which the excess is at least the above expected value. (d) We use the costructio i the hit. First, let s check that E( S ) = 1. Note that if v = (0,..., 0, 0) the Y i = 0 for all i, so S =. Similarly, if v = (0,..., 0, 1) the Y i = 1 for all i, so S =. Now cosider ay other value of v, ad let v j be its leftmost o-zero etry (so j k). For each i {0, 1,..., 1}, let i deote the iteger whose biary expasio differs from that of i oly i positio j. The it is clear that Y i = 1 Y i, ad hece X i = X i. Thus the cotributios to S cacel out i pairs, ad S = 0. Puttig all this together we get E( S ) = 1 2 k+1 ( + ) = 1. To see that the Y i (ad hece the X i ) are -wise idepedet, ote first that, by the same argumet as above, Pr[Y i = 0] = Pr[Y i = 1] = 1 2 for all i. Now fix ay three distict itegers i 1, i 2, i {0, 1,... 1}; we eed to show that Pr[Y i1 = a 1 Y i2 = a 2 Y i = a ] = 1 8 for ay triple of values a j {0, 1}. The sample poits (vectors v) correspodig to this triple are solutios to the equatios Bv T = a T, where a = (a 1, a 2, a ), ad B is the (k + 1) matrix whose rows are b i1, b i2 ad b i. We claim that the rows of B are liearly idepedet over GF [2]; for if ot the some combiatio of them would have to sum to 0, ad sice all have 1 i the last positio this meas that some pair must sum to 0, which is ot possible sice all three rows are distict. Liear idepedece meas that the system of equatios has full rak, so the dimesio of the set of solutios is k + 1 = k 2, i.e., the umber of solutios is 2 k 2. Sice this is idepedet of the triple a, all eight triples must have the same probability, 1 8.