On the Intrinsic Value of Information Amos Golan Department of Economics Director of Info-Metrics Institute American University Radu Balan Department of Mathematics, Center for Scientific Computation and Mathematical Modeling, Norbert Wiener Center University of Maryland 31 October 2014 Recent Innovations in Info-Metrics American University , Washington D.C. Acknowledgments Thanks to sponsor: 2 1. Motivation What the problem is NOT about: • Information: Choose your own preferred definition of information (entropy): – Shannon’s: πΌ π = − π π=1 ππ log 2 (ππ ) – Hartley’s: π» π = log 2 (π) – Renyi’s: π»πΌ π = • Complexity: 1 log 2 1−πΌ π πΌ π π=1 π – Information/Entropy – Kolmogorov complexity: shortest length of a universal coder 3 What this talk IS about: VALUE of Information • Approach: – We start from a set of axioms to construct value of information functions – We investigate existence and uniqueness of such functions – We parameterize various classes of value of information functions • Ingredients that may be added at a later time: – Partial information – Contradictory information 4 2. Axiomatic Let X denote a finite information set π = π₯1 , π₯2 , … , π₯π Examples: • Chapters in a book • Books in a library • Entries in the Webster’s dictionary • Set of logical propositions in a logical system • List of candidates in an election 5 A value of information function (voif) π£ assigns a number (value) to a subset π΄ of π: π£: π« π → β Definition. A Type A value of information function is a function π£: π« π → β that satisfies: P1. (Normalization) π£ ∅ = 0 , π£ π = 1 ; P2. (Positivity) π£ π΄ ≥ 0 for all π΄ ∈ π«(π) ; P3a. (“Concavity Type A”) (Super-additivity) π£ π΄ ∪ π΅ ≥ π£ π΄ + π£ π΅ − π£(π΄ ∩ π΅) for every π΄, π΅ ∈ π«(π). We denote by β±π΄ (π) the set of Type A v.o.i.f.’s. Note β±π΄ π ⊂ 0,1 2π ⊂β 2π 6 Comments: • P1+P2: nonnegative and normalized to [0,1] • Reasonable assumptions provided no contradictory information, and that X provides full information. • “Super-additivity”: – For disjoint sets π΄ ∩ π΅ = ∅, π£ π΄ ∪ π΅ ≥ π£ π΄ + π£(π΅) (*) Formalizes the statement “value of a system is larger than the sum of values of its parts” – For non-disjoint sets, e.g. π΄ = π΅, (*) cannot be extended trivially to: π£ π΄ = π£ π΄ ∪ π΄ ≥ π£ π΄ + π£ π΄ = 2π£(π΄) The formal axiom P3a removes the value of intersection. 7 Axiomatic – Part 2 The concavity hypothesis admits a different form: Definition. A Type B Value of Information Function is a function π£: π« π → β that satisfies: P1. (Normalization) π£ ∅ = 0 , π£ π = 1 ; P2. (Positivity) π£ π΄ ≥ 0 for every π΄ ∈ π«(π) ; P3b. (“Concavity Type B”): π£ π΄ ∪ π΅ ≥ π£ π΄ + π£(π΅) for every π΄, π΅ ∈ π«(π) so that π΄ ⊄ π΅ and π΅ ⊄ π΄. We denote by β±π΅ (π) the set of Type B v.o.i.f.’s. Note β±π΅ π ⊂ 0,1 2π ⊂ π 2 β 8 Discussion of P3b: P3b is a different generalization of the inequality (*) , which is P3a for disjoint sets. In general: • P3a implies a growth at least linearly in terms of subset size : π£ π΄ ≥ π|π΄| • P3b implies a growth at least exponentially in terms of subset size: π£ π΄ ≥ π1 π π2 |π΄| . 9 Type A VOIF: Extreme Points (1) We seek extreme points (vertices) of the convex set β±A (X). “Dictatorial” Extreme Points In fact for any ∅ ≠ π΄ ⊂ π, the following type A VOIF 1 ππ π΄⊂π΅ ππ΄ : π« π → β , ππ΄ π΅ = 0 ππ ππ‘βπππ€ππ π is an extreme point in β±π΄ (π). Explicitly: if ππ΄ = π‘π£1 + 1 − π‘ π£2 for some 0 < π‘ < 1 and π£1 , π£2 type A voif’s then π£1 = π£2 = ππ΄ . 10 Type A VOIF: Extreme Points (2) Another class of extreme points in β±π΄ π : “Democratic” (Equalitarian) Extreme Points For every 2 ≤ π ≤ π − 1, the following type A VOIF π π π΄ = π΄ −π+1 ππ π−π+1 0 ππ π΄ ≥π π΄ <π is an extreme point. 11 Type B VOIF: Extreme Points (1) The situation is slightly more complicated. So far we identified the following classes of extreme points: The “dictatorial” voif’s: For each subset π΄ ∈ π«(π), Ψπ΄,1 2 π΅ −π π΅ = 0 ππ ππ π΄βπ΅ ππ‘βπππ€ππ π A second type of dictatorial voif: Ψπ΄,2 2 π΅ −π π΅ = 2 π΄ −π+1 0 ππ ππ ππ π΄βπ΅ π΄=π΅ ππ‘βπππ€ππ π 12 Type B VOIF: Extreme Points (2) The “democratic” (“equalitarian”) Type B VOIF extreme points are given by: Ψ(π) 2 π΄ −π π΄ = 0 ππ ππ π΄ ≥π π΄ <π where 2 ≤ π ≤ π. Note Ψ (π) factors through the cardinal function. 13 4. An Application: List Rankings Consider π = πΆπππππππ‘π1 , … , πΆπππππππ‘ππ and let π£: π« π → β represent an assignment of preferences. Assume v is a Type B VOIF (suitable for “game changer”): π£ π΄ = max π£ π΄ β π + π£(π΄ β π ) , π΄ ≥2 π,π∈π΄ π≠π 1 Assume π£ πΆπππππππ‘ππ = ππ for 1 ≤ π ≤ π, where each π (π1 , … , ππ ) is drawn randomly from a fixed distribution, and π is the normalization factor so that π£ π = 1. Problem: How many distinct rankings can be achieved on π«(π) and what are their probability distributions. 14 List Rankins: Results We consider the case of uniform distribution for individual preferences π1 , … , ππ . Results over 1,000,000 simulations: π =π 2 3 4 5 6 7 4 8 16 32 64 128 #πππ π ππππ π πππππππ = π« π −2 ! = 2π − 2 ! 2 720 8.717 1010 2.652 1032 3.147 1085 2.372 10211 # π΄πβπππ£ππ π πππππππ 2 12 480 197379 ? ? π« π = 2π 15 List Rankings: Results (2) N = 3 , 2N = 8 , (2N-2)!= 720 , Achieved=12 16 List Rankings: Results (3) N = 4, 2N = 16 , (2N-2)!= 8.717 1010, Achieved=480 17 List Rankings: Results (4) N = 5 , 2N = 32 , (2N-2)!= 2.652 1032 , Achieved=197379 18 5. Conclusions • In this talk we studied the class of concave (super additive) nonnegative normalized functions quantifying value of information. Motivation: Value of Information. • We show their set form a convex set. • We identified classes of extreme points (vertices). • We presented an application on list rankings. 19 Thank you! 20