Back to composing work packagesAll experiments in this enquiry

Project wiring diagrams · Essai 07

Can the boundary win
under every dependence?

Knowing each package's reliability does not tell us how their failures coincide. Compare two assembly boundaries against every joint distribution compatible with what is actually specified.

Essai 6 compared independent and common exposure. Those are two possible couplings. This essai computes sharp bounds over all compatible couplings and exhibits distributions attaining both ends. It can certify dominance, a tie, ambiguity, or inconsistent assumptions.

1. Hold the packages fixed. Change which outputs depend on each other.

Three fictional packages A, B and C each return success/failure status. Each design delivers two service outputs. One output requires a pair of packages; the other stands alone. Both outputs count equally. This deliberately stipulated acceptance model isolates a boundary choice; it is not a structural design for a real wall or shelter.

Left boundary

A + B → ANDandC alone

Accepted output count: A×B + C.

Right boundary

A aloneandB + C → AND

Accepted output count: A + B×C.

Success-status information is used in counterfactual comparisons; no crew, material or physical package is copied. The acyclic gate diagram supplies a deterministic interpretation of each assembly. Probability is added only by the joint distribution over the eight input states.

2. Specify the known probabilities and the acceptance threshold

A joint all-failure probability is partial dependence information. It does not identify a causal mechanism, and unspecified does not mean zero. All numbers are assumptions. Controls are ephemeral; nothing is saved or sent.

3. Inspect a distribution that attains a bound

A B C statusProbabilityLeft accepted countRight accepted countThreshold met

Why no other distribution can beat this bound

The listed weighted constraint sum bounds the objective separately in every one of the eight states. Averaging those inequalities over any feasible probability distribution preserves the inequality. The displayed witness attains equality, proving the bound sharp for this finite model. Both the distribution and the inequality coefficients are independently checked.

Which observations would settle the choice?

For one required output, compare the success of A and C on the same occasions when B fails. The left diagram’s lead equals P(B fails) × [P(C succeeds given B fails) − P(A succeeds given B fails)]. When B always succeeds the lead is zero. Estimating the full joint distribution is unnecessary for this particular comparison.

The result is about the actual threshold

At the starting equal A/C marginals, either boundary can win depending on dependence. The difference is confined to states where B fails and exactly one of A and C succeeds. Marginals alone do not fix how much probability lies in each of those states.

Require both outputs, and the two events become exactly the same: all three packages must succeed. The relative bound collapses to zero for every feasible joint distribution. This is a structural equality, not a numerical coincidence or an assumption of independence.

Crucially, the comparison optimises Left minus Right under the same distribution. Subtracting separately optimised probability bounds would allow two incompatible worlds of dependence and can give an unnecessarily weak comparison.

The finite optimisation and its proof obligations
One probability pₛ for each s∈{0,1}³.
pₛ≥0; Σpₛ=1; Σs_A pₛ=q_A; Σs_B pₛ=q_B; Σs_C pₛ=q_C.
Optional: p₀₀₀=q_joint.
Minimise / maximise Σpₛ · [1{Left(s)≥k} − 1{Right(s)≥k}].

This is a bounded linear programme on a finite probability polytope. The code enumerates every basis of the equality constraints, solves the associated linear system, and retains feasible vertices. A linear objective attains an optimum at a vertex. More importantly, every displayed extremum also includes a dual inequality valid in all eight states; a feasible attaining distribution plus that inequality certifies global optimality independently of the enumeration method.

There are four independent equality rows, or five with the joint failure constraint. Floating-point arithmetic is used in the page with explicit numerical tolerances. The exposed finite parameter vocabulary is additionally checked using exact rational certificate arithmetic. Inconsistent constraints produce no conclusion and no invented witness.

Source → concept → construction → limit

Patterson–Spivak–Vagner, pp.1–4, distinguishes acyclic wiring syntax from the algebra interpreting it. Here ordinary Boolean/count functions interpret the acceptance gates; category theory does not supply a probability law.

Embrechts–Puccetti, Bounds for functions of dependent risks (2006), §§1–3 and §5.1, supplies the fixed-marginal dependence problem and discrete linear-programming method. Their warning that a proposed higher-dimensional bound need not be sharp matters here: sharpness is established by this finite model's own attaining distribution and dual certificate. The shelter interpretation, two-output functions and comparison are our construction.

No dependence estimate, causal risk model, calibrated probability or engineering suitability is supplied. The model holds package marginal reliability fixed across rewiring; a real boundary change could alter those marginals or change dependence even when the marginals stay fixed. It omits cost, schedule and the value of unequal outputs. Those are questions to add explicitly, not benefits implied by a better acceptance bound.

Return to Essai 6: the same package, a different assembly