CS 201: Adversary arguments. This handout presents two lower bounds for selection problems using adversary arguments ëknu73,

Similar documents
x y z HD(x, y) + HD(y, z) HD(x, z)

Counting on r-fibonacci Numbers

Logarithms APPENDIX IV. 265 Appendix

arxiv: v2 [math.co] 15 Oct 2018

Ch 9 Sequences, Series, and Probability

PERMUTATIONS AND COMBINATIONS

Permutation Enumeration

A New Space-Repetition Code Based on One Bit Feedback Compared to Alamouti Space-Time Code

COMBINATORICS 2. Recall, in the previous lesson, we looked at Taxicabs machines, which always took the shortest path home

THE LUCAS TRIANGLE RECOUNTED. Arthur T. Benjamin Dept. of Mathematics, Harvey Mudd College, Claremont, CA Introduction

PERMUTATIONS AND COMBINATIONS

Unit 5: Estimating with Confidence

Application of Improved Genetic Algorithm to Two-side Assembly Line Balancing

Lecture 4: Frequency Reuse Concepts

X-Bar and S-Squared Charts

}, how many different strings of length n 1 exist? }, how many different strings of length n 2 exist that contain at least one a 1

Name Class. Date Section. Test Form A Chapter Chapter 9 Infinite Series. 1 n 1 2 n 3n 1, n 1, 2, 3, Find the fourth term of the sequence

8. Combinatorial Structures

Chapter (6) Discrete Probability Distributions Examples

Sorting, Selection, and Routing on the Array with Reconfigurable Optical Buses

Roberto s Notes on Infinite Series Chapter 1: Series Section 2. Infinite series

Optimal Arrangement of Buoys Observable by Means of Radar

On Parity based Divide and Conquer Recursive Functions

A SELECTIVE POINTER FORWARDING STRATEGY FOR LOCATION TRACKING IN PERSONAL COMMUNICATION SYSTEMS

Grade 6 Math Review Unit 3(Chapter 1) Answer Key

PROJECT #2 GENERIC ROBOT SIMULATOR

7. Counting Measure. Definitions and Basic Properties

AC : USING ELLIPTIC INTEGRALS AND FUNCTIONS TO STUDY LARGE-AMPLITUDE OSCILLATIONS OF A PENDULUM

Intermediate Information Structures

A study on the efficient compression algorithm of the voice/data integrated multiplexer

lecture notes September 2, Sequential Choice

2. There are n letter and n addressed envelopes. The probability that all the letters are not kept in the right envelope, is. (c)

Procedia - Social and Behavioral Sciences 128 ( 2014 ) EPC-TKS 2013

POWERS OF 3RD ORDER MAGIC SQUARES

13 Legislative Bargaining

Counting and Probability CMSC 250

On the Number of Permutations on n Objects with. greatest cycle length

Energy-Optimal Online Algorithms for Broadcasting in Wireless Networks

[MT93a, ALS94, Nie95, MNR95]. All these algorithms exploit kow characterizatios of extesios of default theories i terms of sets of geeratig defaults,

APPLICATION NOTE UNDERSTANDING EFFECTIVE BITS

Combinatorics. Chapter Permutations. Reading questions. Counting Problems. Counting Technique: The Product Rule

Counting III. Today we ll briefly review some facts you dervied in recitation on Friday and then turn to some applications of counting.

Laboratory Exercise 3: Dynamic System Response Laboratory Handout AME 250: Fundamentals of Measurements and Data Analysis

CHAPTER 5 A NEAR-LOSSLESS RUN-LENGTH CODER

PERMUTATION AND COMBINATION

HOW BAD RECEIVER COORDINATES CAN AFFECT GPS TIMING

We often find the probability of an event by counting the number of elements in a simple sample space.

Fingerprint Classification Based on Directional Image Constructed Using Wavelet Transform Domains

Using Color Histograms to Recognize People in Real Time Visual Surveillance

General Model :Algorithms in the Real World. Applications. Block Codes

Discrete Mathematics and Probability Theory Spring 2014 Anant Sahai Note 12

1. How many possible ways are there to form five-letter words using only the letters A H? How many such words consist of five distinct letters?

A generalization of Eulerian numbers via rook placements

A Math Learning Center publication adapted and arranged by. EUGENE MAIER and LARRY LINNEN

Test Time Minimization for Hybrid BIST with Test Pattern Broadcasting

BOUNDS FOR OUT DEGREE EQUITABLE DOMINATION NUMBERS IN GRAPHS

The Institute of Chartered Accountants of Sri Lanka

(2) The MOSFET. Review of. Learning Outcome. (Metal-Oxide-Semiconductor Field Effect Transistor) 2.0) Field Effect Transistor (FET)

AMC AMS AMR ACS ACR ASR MSR MCR MCS CRS

The Fundamental Capacity-Delay Tradeoff in Large Mobile Ad Hoc Networks

15 min/ Fall in New England

CS3203 #5. 6/9/04 Janak J Parekh

WAVE-BASED TRANSIENT ANALYSIS USING BLOCK NEWTON-JACOBI

A New Basic Unit for Cascaded Multilevel Inverters with the Capability of Reducing the Number of Switches

Article Modified Adversarial Hierarchical Task Network Planning in Real-Time Strategy Games

ON THE FUNDAMENTAL RELATIONSHIP BETWEEN THE ACHIEVABLE CAPACITY AND DELAY IN MOBILE WIRELESS NETWORKS

Introduction to Wireless Communication Systems ECE 476/ECE 501C/CS 513 Winter 2003

Data Acquisition System for Electric Vehicle s Driving Motor Test Bench Based on VC++ *

GENERATE AND MEASURE STANDING SOUND WAVES IN KUNDT S TUBE.

Density Slicing Reference Manual

Technical Explanation for Counters

Join a Professional Association

Design of FPGA Based SPWM Single Phase Inverter

4. INTERSYMBOL INTERFERENCE

Compound Controller for DC Motor Servo System Based on Inner-Loop Extended State Observer

MEI Core 2. Logarithms and exponentials. Section 2: Modelling curves using logarithms. Modelling curves of the form y kx

Final exam PS 30 December 2009

Zonerich AB-T88. MINI Thermal Printer COMMAND SPECIFICATION. Zonerich Computer Equipments Co.,Ltd MANUAL REVISION EN 1.

ASample of an XML stream is:

Lecture 29: MOSFET Small-Signal Amplifier Examples.

Compression Programs. Compression Outline. Multimedia. Lossless vs. Lossy. Encoding/Decoding. Analysis of Algorithms

3. Error Correcting Codes

A Novel Three Value Logic for Computing Purposes

AkinwaJe, A.T., IbharaJu, F.T. and Arogundade, 0.1'. Department of Computer Sciences University of Agriculture, Abeokuta, Nigeria

CHAPTER 8 JOINT PAPR REDUCTION AND ICI CANCELLATION IN OFDM SYSTEMS

Measurements of the Communications Environment in Medium Voltage Power Distribution Lines for Wide-Band Power Line Communications

Hybrid BIST Optimization for Core-based Systems with Test Pattern Broadcasting

Data Mining of Bayesian Networks to Select Fusion Nodes from Wireless Sensor Networks

AS Exercise A: The multiplication principle. Probability using permutations and combinations. Multiplication principle. Example.

ELEC 350 Electronics I Fall 2014

PHY-MAC dialogue with Multi-Packet Reception

Implementation of Fuzzy Multiple Objective Decision Making Algorithm in a Heterogeneous Mobile Environment

On the Capacity of k-mpr Wireless Networks

AMC AMS AMR ACS ACR ASR MSR MCR MCS CRS

BOTTLENECK BRANCH MARKING FOR NOISE CONSOLIDATION

Modeling and solution for the ship stowage planning problem of coils in the steel industry

A New Design of Log-Periodic Dipole Array (LPDA) Antenna

Time-Space Opportunistic Routing in Wireless Ad Hoc Networks, Algorithms and Performance

A GHz Constant KVCO Low Phase Noise LC-VCO and an Optimized Automatic Frequency Calibrator Applied in PLL Frequency Synthesizer

20. CONFIDENCE INTERVALS FOR THE MEAN, UNKNOWN VARIANCE

Transcription:

CS 01 Schlag Jauary 6, 1999 Witer `99 CS 01: Adversary argumets This hadout presets two lower bouds for selectio problems usig adversary argumets ëku73, HS78, FG76ë. I these proofs a imagiary adversary called a oracle is created to force a algorithm to use the maximum umber of comparisos to solve the problem. The adversary's task is to determie the outcomes of comparisos i a maer which impedes the progress of the algorithm, ad yet is still cosistet. The outcomes of comparisos are cosistet as log as there are ever three keys A, B ad C with AéB, B é Cad C é A. That is, at aytime the comparisos form a partial order o the keys. A selectio algorithm ca be viewed as a touramet i which players participate i a sequece of pairwise matches. Assumig that each player has a distict skill level ad always plays at this level, the outcomes of matches are determied solely by these skill levels. These skill levels are ot kow. Oly by playig a match ca the skill levels of two players be compared. The better player always wis the match ad the outcomes are always cosistet sice they are based o a èhiddeè total orderig of the players. The three quatities of iterest i selectio problems are: 1. V k èè, the umber of comparisos ecessary to determie the k th player out of players,. U k èè, the umber of comparisos ecessary to determie the set cosistig of the top k players out of players, ad 3. W k èè, the umber of comparisos ecessary to determie the 1 st ; d ;:::;k th players out of players. These three quatities are related as follows: U k èè V k èè W k èè: The ærst iequality follows from the observatio that if you have determied through comparisos that x is the k th player the you also must have idetiæed the k, 1 players better tha x as well as the, k players worse tha x. I the ext sectio it is show that W èè =V èè =, +dlog e, ad i the followig sectio it is show that d 3 e, matches is ecessary ad suæciet for ædig the top ad bottom players. More recet results are preseted i ëbj85ë. Fidig the top two players Result 1: The top two players ca be foud i, +dlog e matches. Proof: Cosider the followig algorithm èschedule of matchesè to æd the ærst ad secod players: 1. First hold a sigle elimiatio touramet for the players. That is, pair up the players i b c matches, ad repeat the process with the udefeated players util oe udefeated player, x, remais. 1

. Play a secod sigle elimiatio touramet amog the m players who were defeated by x i the ærst touramet. Sice each player loses at most oce i a sigle elimiatio touramet ad at the ed oly oe player is udefeated, there are exactly, 1 losses ad hece, 1 matches. So the ærst touramet has exactly, 1 matches. Likewise, the umber of matches i the secod touramet is m, 1 where m is the umber of players who participated, lost to x i the ærst touramet. To determie m, cosider the umber of rouds played i the ærst touramet. I each roud of a sigle elimiatio tourmaet the players are paired ad d e players remai for the ext roud. Hece there will be at most dlog e rouds. I each roud x will defeat oe player, so m = dlog e. Summig the umber of matches played i the two touramets we obtai,, +dlog e: Result : Ay algorithm for ædig the ærst ad secod players will require at least, +dlog e matches. Proof: Suppose we have a arbitrary algorithm A for ædig the ærst ad secod player. Sice A is completely arbitrary we ca assume othig about the matches that A arrages. We ivet a æctitious adversary for A called a oracle as follows. As the matches orgaized by A are played our oracle keeps track of the followig iformatio: 1. let TOP = fcurretly udefeated playersg,. for ay player x TOP, let Domèx; 0è = fxg 3. for ay player x TOP ad ié0, let Domèx; iè=fplayers whose ærst defeat was to a player i Domèx; i, 1è g, 4. let Domèxè = S 1 i=0 Domèx; iè, ad 5. let Wièxè = jdomèx; 1èj. Our oracle uses the followig rules to determie the outcome of matches: 1. I a match betwee two players curretly i TOP, the player with the larger Wièè value will wi. If the two players have the same Wièè value the the wier is selected arbitrarily.. I a match betwee a player curretly i TOP ad a player ot i TOP, the player i TOP wis. 3. If either player is curretly i TOP, the the wier of the match is selected arbitrarily i a maer cosistet with the outcome of previous matches. The outcomes of the matches determied by our oracle are always cosistet sice the wier is always a udefeated player i the ærst two rules ad the outcome i rule 3 is selected to esure cosistecy. Iitially each player y belogs to Domèyè ad TOP. Whe y suæers its ærst loss, the oracle's secod rule esures that y loses to player z which is also i TOP. Soy alog with all of the other players i Domèyè joi Domèzè. Hece at aytime durig the matches each player belogs to Domèzè for some player z curretly i TOP. Whe the algorithm termiates there is oly oe player i TOP, say x, ad all of the players are i Domèxè.

Claim: After k matches ay player x still i TOP satisæes, jdomèxèj Wièxè : è1è Proof of Claim: By iductio o the umber of matches, k. Base: k = 0. No matches have bee played so jdomèxèj =1= 0 = Wièxè for all players. Step: Assume the claim is true after the ærst k, 1 matches ad cosider the k th match ad a player x still i TOP after this match. Note that sice x is still udefeated after the k th match, x either wo or did ot play ithek th match. Case 1 x did ot play i the k th match. I this case, we had jdomèxèj Wièxè before the match. Wièxè does't chage sice x does't play. Is it possible for Domèxè tochage? The aswer is o sice a player would joi Domèxè oly if it is curretly i TOP ad loses to a player i Domèxè èother tha x i this caseè. However, such a loss is ot possible accordig to the oracle's secod rule. Sice either Domèxè or Wièxè chages, Equatio è1è still holds after the match. Case x played y who was ot i TOP before the k th match. Sice y was ot i TOP, agai either Domèxè or Wièxè chages sice this is ot y's ærst defeat ad Equatio è1è still holds after the match. Case 3 x played y who was i TOP before the k th match. Accordig to the oracle's ærst rule, we have Wièxè Wièyè sice x wo. After the match Domèxè will become Domèxè ë Domèyè ad we have, jdomèxè ë Domèyèj jdomèxèj + jdomèyèj Wièxè + Wièyè æ Wièxè = Wièxè+1 : Sice after the match Wièxè will go up by 1, the Equatio è1è will still hold. Whe algorithm A termiates there will be oly oe player i TOP, say x. Sice all players will be i Domèxè, we have Wièxè jdomèxèj = ; or Wièxè dlog e: Cosider the players i Domèx; 1è. There are Wièxè of them. Each of them suæered their ærst defeat to x. Ayoe of these players that does ot lose a secod match, could be the secod best player. Algorithm A caot rule them out as secod best players, sice they have oly bee show to be worse tha x. I order to determie the secod best player, algorithm A must have arraged matches i which all but oe of the players i Domèx; 1è lost a secod time. The umber of matches played correspods exactly to the umber of losses. There are, 1 ærst losses ad at least Wièxè, 1 secod losses, so the umber of matches used by A is at least, +dlog e: 3

Fidig the top ad bottom players Result 3: The top ad bottom players ca be foud i d 3 e, matches. Proof: Cosider the followig algorithm èschedule of matchesè to æd the ærst ad last players: 1. First hold a sigle elimiatio touramet for the players. That is, pair up the players i b c matches, ad repeat the process with the udefeated players util oe udefeated player, x, remais.. Collect the d e players who did ot wi i the ærst roud of the previous touramet. èif is odd this will iclude the player who did ot play i the ærst roud.è Play a touramet with these players i which the loser advaces to the ext roud rather tha the wier. That is, pair up the players i matches, ad repeat the process with the wiless players util oe wiless player, z, remais. After the ærst touramet all but oe player is udefeated so the top player has bee determied i, 1 matches. The oly players who have ot wo ay matches are those that lost or possibly did ot play i the ærst roud. All the other players have wo at least oe match ad therefore caot possibly be the worst player. So to determie the worst player, matches are played amog the at most d e players who are still wiless. After the secod touramet all but oe player, z, will have wo a match. Oly z will be wiless ad so z is the worst player. The secod touramet requires d e,1 matches so altogether, a total of d 3 e, matches are played. Result 4: Fidig the top ad bottom players requires d 3 e, matches. Proof: As before we assume a arbitrary algorithm A for ædig the ærst ad last players ad we ivet a æctitious adversary for A called a oracle. Durig the sequece of matches arraged by A we keep track of the followig types of players. U is the group of players that have ever played; they are utested. W is the group of players that have played ad ever lost; they are wiers. L is the group of players that have played ad ever wo; they are losers. B is the group of players that have both wo ad lost matches; they are both losers ad wiers. After m matches we will represet the state of the touramet byèu; w; l; bè where u, w, l, ad b are the umbers of players i U, W, L, ad B respectively. The iitial state before ay matches are played is è; 0; 0; 0è ad at the ed, whe the top ad bottom players are kow, the state must be è0; 1; 1;, è. How the state chages after a match depeds o the group memberships of the wier ad loser; the state trasistios are give i Table 1. If wier If loser is i is i U W L B U èu, ;w+1;l+1;bè èu, 1;w;l;b+1è èu, 1;w+1;l;bè èu, 1;w+1;l;bè W èu, 1;w;l+1;bè èu; w, 1;l;b+1è èu; w; l; bè èu; w; l; bè L èu, 1;w;l;b+1è èu; w, 1;l, 1;b+è èu; w; l, 1;b+1è èu; w; l, 1;b+1è B èu, 1;w;l+1;bè èu; w, 1;l;b+1è èu; w; l; bè èu; w; l; bè Table 1: If the curret state is èu; w; l; bè the after a match the ext state ca be foud i the row ad colum correspodig to the groups èbefore the matchè cotaiig the match's wier ad loser. Our oracle tries to impede the progress of the algorithm by miimizig the chage i state. It decides outcomes of matches as follows: 4

1. A player i W always wis agaist a player ot i W.. A player i L always loses agaist a player ot i L. 3. Otherwise the wier is selected arbitrarily i a maer cosistet with the outcome of previous matches. The outcome is cosistet sice players i L ca always lose without violatig cosistecy ad players i W ca always wi without violatig cosistecy. I Table the outcomes prohibited by the oracle have bee oted. All players start i U,, must move tob while oe player moves to W ad If wier If loser is i group is i group U W L B U èu, ;w+1;l+1;bè ot possible èu, 1;w+1;l;bè èu, 1;w+1;l;bè W èu, 1;w;l+1;bè èu; w, 1;l;b+1è èu; w; l; bè èu; w; l; bè L ot possible ot possible èu; w; l, 1; b + 1è ot possible B èu, 1;w;l + 1;bè ot possible èu; w; l; bè èu; w; l; bè Table : Table 1 modiæed to reæect state chages which the oracle will ot allow. oe to L. We ca view each match as trasferrig uits from u to b, leavig oe uit behid i both w ad l. I the remaiig possible state chages, there is o match which trasfers uits from u to b, so trasferrig a uit from u to b must be accomplished by trasferrig it ærst to w or l ad the to b. Hece a total è, è + =, uit trasfers must occur. All matches trasfer at most oe uit with the exceptio of matches betwee two players i U, but at most b c matches of this type ca be played. The remaiig,, b c uits would have to be trasferred oe at a time. Thus the miimum umber of matches ecessary to accomplish all the trasfers is 3 +,, =,, =, : Refereces ëbj85ë ëfg76ë ëhs78ë Samuel W. Bet ad Joh W. Joh. Fidig the media requires comparisos. I Proceedigs of the 17 th ACM Symposium o Theory of Computig, pages 13í16, 1985. Frak Fusseegger ad Harold N. Gabow. Usig compariso trees to derive lower bouds for selectio problems. I Proceedigs of the 17 th IEEE Symposium o Foudatios of Computer Sciece, pages 178í18, 1976. Ellis Horowitz ad Sartaj Sahi. Fudametals of Computer Algorithms. Computer Sciece Press, Rockville, Marylad, 1978. ëku73ë Doald E. Kuth. The Art of Computer Programmig: Sortig ad Searchig. Addiso Wesley, Readig, Massachusetts, 1973. 5