Information Theory, Inference, and Learning ... - Inference Group

Copyright Cambridge University Press 2003. On-screen viewing permitted. Printing not permitted. http://www.cambridge.org/0521642981You can buy this book for 30 pounds or $50. See http://www.inference.phy.cam.ac.uk/mackay/itila/ for links.102 5 — Symbol Codes5.7 SummaryKraft inequality. If a code is uniquely decodeable its lengths must satisfy∑2 −l i≤ 1. (5.25)iFor any lengths satisfying the Kraft inequality, there exists a prefix codewith those lengths.Optimal source codelengths for an ensemble are equal to the Shannoninformation contentsl i = log 21p i, (5.26)and conversely, any choice of codelengths defines implicit probabilitiesq i = 2−l iz . (5.27)The relative entropy D KL (p||q) measures how many bits per symbol arewasted by using a code whose implicit probabilities are q, when theensemble’s true probability distribution is p.Source coding theorem for symbol codes. For an ensemble X, there existsa prefix code whose expected length satisfiesH(X) ≤ L(C, X) < H(X) + 1. (5.28)The Huffman coding algorithm generates an optimal symbol code iteratively.At each iteration, the two least probable symbols are combined.5.8 Exercises⊲ Exercise 5.19. [2 ] Is the code {00, 11, 0101, 111, 1010, 100100, 0110} uniquelydecodeable?⊲ Exercise 5.20. [2 ] Is the ternary code {00, 012, 0110, 0112, 100, 201, 212, 22}uniquely decodeable?Exercise 5.21. [3, p.106] Make Huffman codes for X 2 , X 3 and X 4 where A X ={0, 1} and P X = {0.9, 0.1}. Compute their expected lengths and comparethem with the entropies H(X 2 ), H(X 3 ) and H(X 4 ).Repeat this exercise for X 2 and X 4 where P X = {0.6, 0.4}.Exercise 5.22. [2, p.106] Find a probability distribution {p 1 , p 2 , p 3 , p 4 } such thatthere are two optimal codes that assign different lengths {l i } to the foursymbols.Exercise 5.23. [3 ] (Continuation of exercise 5.22.) Assume that the four probabilities{p 1 , p 2 , p 3 , p 4 } are ordered such that p 1 ≥ p 2 ≥ p 3 ≥ p 4 ≥ 0. LetQ be the set of all probability vectors p such that there are two optimalcodes with different lengths. Give a complete description of Q. Findthree probability vectors q (1) , q (2) , q (3) , which are the convex hull of Q,i.e., such that any p ∈ Q can be written aswhere {µ i } are positive.p = µ 1 q (1) + µ 2 q (2) + µ 3 q (3) , (5.29)