DCT Lexicon · Robustness
Kullback–Leibler divergence
KL(Q ∥ P)
\mathrmKL(Q \Vert P) = \mathbbE_Q \log(dQ/dP): how far a law Q reweights the futures of P.
Vol. II, Glossary of Key Terms, p. 399, its first sentence
In depth
In the book
, how far a law reweights the futures of . A ball of radius keeps a chance exactly when the law keeps the smaller with ; its dual is the entropic criterion (Theorem 11.2).
Vol. II, Glossary of Key Terms, p. 399, verbatim
Formal statement
Let , . (i) Dual form. For a bounded loss , (ii) Worst chance. For an event of chance under , the worst chance over the ball is the with , or 1 if .
Vol. II, Theorem 11.2(i)–(ii) (Divergence Families), verbatim
Executive example
Over a Kullback–Leibler ball of radius d around Chapter 9’s law: the worst chance of a breach of Volume I’s package, 5 percent under the law, and the chance the law must keep for the ball to keep 5 percent; dashed, the radii of the records’ doubts (closures 0.44, spread 0.60, both 1.04).
Vol. II, Figure 11.1 caption, verbatim
Common misreading
A common mistake is to size a ball by intuition rather than by the doubts it must hold.
Vol. II, LOS 11.2, verbatim
How AXIOM uses it
On the Robust vs optimal tab, for each Kullback–Leibler radius AXIOM computes the appetite the plan must keep under the declared law for the ball to keep the board's appetite, and the plan Optimize finds there. No radius is assumed: the curve is shown across radii.
Drafted for the author's approval, from AXIOM's method passport "Robust plan and the cost of doubt" (D-063)
In the book: Vol. II, Ch. 11: Distributionally Robust Enterprise Optimization
Related: Wasserstein Distributionally Robust Optimization · Price of robustness (cost of doubt) · Robust optimization (max-min plan)
V2.Thm.11.2 · V2.Def.11.1 · V2.WE.11.1


