Introduction to Information Theory. Part 2

Transcription

1 Introduction to Information Theory Part 2 1

2 A General Communication System CHANNEL Information Source Transmitter Channel Receiver Destination 2

3 Information: Definition Information is quantified using probabilities. Given a finite set of possible messages, associate a probability with each message. A message with low probability represents more information than one with high probability. Definition of Information: Where is the probability of the message Base 2 is used for the logarithm so is measured in bits Trits for base 3, nats for base, Hartleys for base 10 3

4 Some properties of Information is non negative. Information we get from observing two independent events occurring is the sum of two information(s). is monotonic and continuous in Slight changes in probability incur slight changes in information. We get zero information from an event whose probability is 1. 4

5 Example: Information in a coin flip 5

6 Independent Events: 2 Coin flips There are four possibilities: HH, HT, TH, TT i.e. Additive property: 6

7 Example: Text Analysis a b c d e f g h i j k l m n o p q r s t u v w x y z SPC SPC SPC e e t t a o a o i n s r i n s r h h l l d c u m f p g w y b v k x q j z d c u m f p g w y b v k x q j z 7

8 Example: Text Analysis Letter Freq. I a b c d e f g h i j k l m n o p q r s t u v w x y z SPC

9 Definition of Entropy Information ( ) is associated with known events/messages Entropy ( ) is the average information w.r.to all possible outcomes. Given, Characterizes an information source. 9

10 Example: A 3 event Source ) 10

11 Example: Text Analysis Letter Freq. I a b c d e f g h i j k l m n o p q r s t u v w x y z SPC Aka, First Order Entropy. 11

12 Entropy (2 outcomes) 12

13 Entropy: Properties Entropy is maximized if is uniform. Additive property for independent events. If and are not independent. 13

14 Entropy of things Entropy of English text is approx 1.5 bits Entropy of the human genome <= 2 bits Entropy of a black hole is ¼ of the area of the outer event horizon. Value of information in economics is defined in terms of entropy. E.g. Scarcity 14

15 Entropy: What about it? Does have a maximum? Where? Is entropy a good name for this stuff? How is it related to entropy in thermodynamics? How does entropy help in communication? What else can we do with it? Why use the letter? 15

16 Entropy & Codes Entropy is closely related to the design of efficient codes for random sources. Provides foundations for techniques of compression, data search, encryption, correction of communication errors, etc. Essential to the study of life sciences, economics, etc. 16

17 Coding: Basics Events of an information source:,, A code is made up of codewords from a code alphabet (e.g. {0, 1}, {., }, etc.) A code is an assignment of codewords to source symbols. 17

18 Coding: Basics Block code: When all codes have the same length. For example, ASCII (8 bits) Average Word Length: More generally, 1 A code is efficient if it has the smallest average word length. (Turns out entropy is the benchmark ) 18

19 Coding: Basics Singular (not unique) codes Nonsingular (unique) codes Symbol Singular Code Nonsingular Code A 00 0 B C D

20 Coding: Basics Singular (not unique) codes Nonsingular (unique) codes instantaneous codes (every word can be decoded as soon as it is received) Symbol Singular Code Nonsingular Code A 00 0 B C D Not an instantaneous Code! 20

21 Example: Avg. Code Length (L) Symbol p Codeword A B C D E

22 Example: Source Entropy (H) Symbol p Codeword A B C D E log log log

23 Example: & Symbol p Codeword A B C D E log Is there a relationship between and? log log

24 Average Code Length & Entropy Average length bounds: Grouping symbols together: 24

25 Average Code Length & Entropy Average length bounds: Grouping symbols together: 25

26 Average Code Length & Entropy Average length bounds: Grouping symbols together: 26

27 Average Code Length & Entropy Average length bounds: Grouping symbols together: 27

28 Average Code Length & Entropy Average length bounds: 1 Grouping symbols together: lim H 28

29 Shannon s First Theorem By coding sequences of independent symbols (in ), it is possible to construct codes such that The price paid for such improvement is increased coding complexity (due to increased ) and increased delay in coding. 29

30 Entropy & Coding: Central Ideas Use short codes for highly likely events. This shortens the average length of coded messages. Code several events at a time. Provides greater flexibility in code design. 30

31 Data Compression: Huffman Coding A 0.3 B 0.2 C 0.2 D 0.2 E

32 Huffman Coding: Reduction Phase A B C D E

33 Huffman Coding: SplittingPhase A B C D E

34 Huffman Coding: SplittingPhase A B C D E

35 Huffman Codes Nonsingular Instantaneous Efficient Non unique Powers of a source lead closer to Requires knowledge of symbol probabilities 35

36 Design Huffman Codes 36

37 References Eugene Chiu, Jocelyn Lin, Brok Mcferron, Noshirwan Petigara, Satwiksai Seshasai: Mathematical Theory of Claude Shannon: A study of the style and context of his work up to the genesis of information theory. MIT 6.933J / STS.420J The Structure of Engineering Revolutions Luciano Floridi, 2010: Information: A Very Short Introduction, Oxford University Press, Luciano Floridi, 2011: The Philosophy of Information, Oxford University Press, James Gleick, 2011: The Information: A History, A Theory, A Flood, Pantheon Books, Zhandong Liu, Santosh S Venkatesh and Carlo C Maley, 2008: Sequence space coverage, entropy of genomes and the potential to detect non human DNA in human samples, BMC Genomics 2008, 9:509 David Luenberger, 2006: Information Science, Princeton University Press, David J.C. MacKay, 2003: Information Theory, Inference, and Learning Algorithms, Cambridge University Press, Claude Shannon & Warren Weaver, 1949: The Mathematical Theory of Communication, University of Illinois Press, W. N. Francis and H. Kucera: Brown University Standard Corpus of Present Day American English, Brown University,

38 38

39 Additive property.,, log log 39