INFORMATION THEORY

Syllabus (every class = 2 hours)

CLASS 1:
Intro:
Course organization, objectives, textbooks, exam details. Sneaky preview of the course, motivations, applications. Assigned Reading of Ch.1 of textbook. Physical justification and definition of entropy. Examples of entropy calculation. Up to sec. 2.1.

CLASS 2:
Definition of joint and conditional entropy, eample 2.2.1. Relative entropy, mutual information and their relation. Chain rules for PMFs and entropy.

CLASS 3:
Relative conditional entropy, conditional mutual information, chain rules for D and I. Inequalities for D and I. max and min of H, H(X|Y)<=H(X) and generalizations. Convex functions. Jensen's inequality, examples.

CLASS 4:
first hour: logsum inequality, convexity of D, concavity of H. concavity of I in p(x) and convexity in p(y|x). Exercise: mixing increases entropy. second hour: Definition of Markov chain and first properties for 3 random variables (RV) X,Y,Z. Data processesng inequality. Counterexample.

CLASS 5:
first hour: sufficient statistics: definition in terms of mutual information. Examples: number of successes in repeated trials; sample mean in estimation of common mean in a vector of independent gaussian RVs. Sufficient statistics and hypothesis testing: factorization theorem. Second hour: Fano inequality. Exercise 2.32.

CLASS 6:
Exercises 2.5, 2.4, 2.27, 2.30 (after brief introduction to the method of Lagrange multipliers), 2.21.

CLASS 7:
Ch 3 asymptotic equipartition property (AEP): introduction. Probability theory refresher: i.p. convergence, Chebychev inequality, Weak law of large numbers, AEP. Typical set and properties. Example with binary sequences.

CLASS 8:
first hour: relation among Typical set and high-probability sets. Theorem 3.3.1. second hour: Problem solving: Exercises 3.8, 3.9. Ch 4: Entropy rates: introduction. Definition of discrete-time stochastic process and stationarity.

CLASS 9:
first hour: introduction to discrete-time Markov chains (DTMC): transition matrix, updatge law (Chapman-Kolmogorov), stationary distribution. Two-state example: state diagram, evolution towards limit distribution. Evaluation with flux balancing. second hour: Entropy rates H and H'. H=H' for stationary processes. Statement of AEP theorem for stationary ergodic sources (Shannon/Breiman/McMillan). Explicit evaluation of H for DTMC. Examples.

CLASS 10:
Doubly-stochastic matrices and uniform steady-state distribution. Connections with entropy as defined in statistical thermodynamics: DTMC on microstates with doubly-stochastic transition matrix. Entropy increases towards steady-state distribution entropy. Example 4 (eq 4.50-4.52). Hidden Markov models (HMM): entropy rate.

CLASS 11:
Problem solving: Ex. 4.1 mixing increases entropy. Conditions for observable Y in a HMM sian una DTMC. examples where Y is not a DTMC. Point a. of Ex. 4.18 on Entropy Rate of stationary but not ergodic process.

CLASS 12:
Problem solving: first hour: points b, c of Ex. 4.18. second hour: Es 4.10 entropy rate of a second order markov process: study of hidden markov chain. Ex. 4.6.

CLASS 13:
Ch 5: Data compression. Examples of codes. Kraft inequality. Search of optimal codes with Lagrange multipliers method. Noiseless coding theorem.

CLASS 14:
Comments on first Shannon Theorem: when p not dyadic. Quasi-optimal Shannon Codes. Shannon super-codes are asymptotically optimal. Extra cost on minimal code length when using a PMF that differs from true PMF. McMillan Theorem: every uniquely decodable theorem satisfies Kraft inequality. Introduction to Huffman codes: examples 1, 2.

CLASS 15:
Huffman codes: example 3 (dummy symbols), Exercise 5.32, example 5.73 (set of different optimal codelengths). Competitive optimality of Shannon code. Proof of optimality of Huffman code.

CLASS 16:
first hour: Optimal compression of Markov sources. Description of Lempel-Ziv algorithm for universal compression. second hour: Channel capacity: introduction, definition of discrete memoryless channel (DM