DCT Lexicon · Optimization
Attribution (one input at a time)
Which declared uncertainty drives the spread of value and the chance of a breach, read by freezing one input at its mean on the same paths while every other keeps its draw: a local reading, not a decomposition.
Drafted for the author's approval from Vol. II, Theorem 4.3 (local sensitivity) and Vol. I, Glossary of Key Terms ("Sensitivity"). AXIOM does not compute Shapley values; the book's Shapley attribution (Vol. II, Definition 13.3) is a different object
In depth
In the book
The derivative of equilibrium or trajectory quantities with respect to parameters, decisions, or shares; computed by the Jacobian calculus and decomposable along propagation walks (Definition 14.6; Theorem 14.4).
Vol. I, Glossary of Key Terms, p. 376 (Sensitivity), verbatim
Not a Shapley value
The only attribution of a prediction that sums to it, treats features symmetrically, gives nothing to a feature that changes nothing, and adds across games (Definition 13.3; Proposition 13.3(iii)); it describes the model under the declared game, interventional or conditional, not causes (Proposition 13.3(iv)).
Vol. II, Glossary of Key Terms, p. 403 (Shapley value), verbatim
Formal statement
One at a time: , where is the P5–P95 width of the uplift (or the joint breach probability) on the same paths. Shapley: The need not add up to the width; the do, by construction.
Drafted for the author's approval: AXIOM's one-at-a-time reading, beside the book's Shapley value (Vol. II, display (13.2))
Common misreading
A common mistake is to read an attribution as a cause.
Vol. II, LOS 13.5, verbatim
How AXIOM uses it
On the Risk factors tab, each random input is frozen at its mean in place, on the same 100,000 paths, and AXIOM reports how much narrower the value's spread becomes and what the joint breach probability is without it. The response says by how much the reductions fail to add up. The reverse-stress multiples sit beside it.
Drafted for the author's approval, from AXIOM's method passport "What drives the spread and the breach" (D-085)
In the book: Vol. II, Ch. 4: Nonlinear and Mixed-Integer Enterprise Optimization · Vol. II, Ch. 13: Machine Learning for Enterprise Transformation · Vol. I, Ch. 12: Enterprise Risk, Resilience, and Robustness Architecture
Related: Reverse stress test · Common random numbers · Tangent Dynamics
V2.Thm.4.3 · V1.WE.12.1 · V2.Def.13.3 · V2.Prop.13.3


