Find your next learning step
A Wasserstein ball can protect capacity decisions from sampling error, but its radius, ground metric and operational meaning must be calibrated and tested together.
Open the chapter contents · Queueing Theory root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Define the ambiguity set
Route 2 · Work the transport ledger
Route 3 · Calibrate and transfer
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE AMBIGUITY SET
1. Why a radius can change a capacity decision
A forecast sample is not the demand distribution itself. A Wasserstein ambiguity set makes that distinction operational: it places a ball of distributions around the empirical sample and asks which capacity survives the worst distribution inside the ball. The method can protect a decision from sampling error, but only if the distance, radius and loss are stated in the same units as the decision.
The central question is not whether a larger radius is “safer”. It is whether the radius represents a defensible amount of distributional uncertainty and whether the resulting conservatism is worth its capacity cost. A robust answer needs calibration, held-out evidence and a record of which probability moves drive the decision.
CHAPTER 2 OF 12 · DEFINE THE AMBIGUITY SET
2. The ambiguity set from first principles
Let P̂ be the empirical distribution of observed demand paths. A Wasserstein ball contains every distribution Q whose transportation distance from P̂ is at most ε. Transporting probability mass over a long ground-metric distance consumes more of the radius than moving it a short distance.
For a capacity action y and loss L(y,ξ), the distributionally robust problem evaluates sup Q E_Q[L(y,ξ)] over that ball. The radius ε is therefore a budget for adversarial probability movement, not an error percentage. Its numerical meaning changes when demand is rescaled, horizons are reweighted or path features are replaced.
CHAPTER 3 OF 12 · WORK THE TRANSPORT LEDGER
3. A three-scenario capacity ledger
Suppose one period has empirical demands 8, 10 and 14 with probabilities 0.50, 0.30 and 0.20. Capacity costs 2 units per place and unmet demand costs 8 units per job. Compare y=10 and y=12.
| Capacity y | Capacity cost | Empirical shortage cost | Empirical total |
|---|---|---|---|
| 10 | 20.0 | 0.20×4×8 = 6.4 | 26.4 |
| 12 | 24.0 | 0.20×2×8 = 3.2 | 27.2 |
The empirical decision is close. That matters because a small, plausible redistribution of mass can reverse it. The arithmetic should be visible before a solver is trusted.
CHAPTER 4 OF 12 · WORK THE TRANSPORT LEDGER
4. A feasible adversarial movement
Use absolute demand difference as the ground metric. Moving 0.10 probability from demand 10 to demand 14 costs 0.10×|14−10|=0.40 units of Wasserstein budget. Under that feasible distribution, demand probabilities become 0.50, 0.20 and 0.30.
The total for y=10 rises to 20+0.30×4×8=29.6. The total for y=12 rises to 24+0.30×2×8=28.8. This one feasible movement proves that y=10 can be exposed within a radius of 0.40; it does not by itself prove that the movement is the exact worst case. The full optimisation must still search the entire ambiguity set.
CHAPTER 5 OF 12 · WORK THE TRANSPORT LEDGER
5. Radius calibration is an evidence problem
A radius can be linked to concentration bounds under stated sampling assumptions, selected by out-of-sample validation, or stress-tested over a policy range. These approaches answer different questions. A theoretical confidence radius concerns coverage of an unknown distribution. A validation radius concerns performance on held-out data. A policy stress radius asks how much misspecification the decision should tolerate.
Fournier and Guillin give non-asymptotic bounds for empirical measures in Wasserstein distance. Their rates depend on dimension and moment conditions. They do not supply a universal queueing radius. Path dimension, dependence and finite sample size must remain visible.
CHAPTER 6 OF 12 · CALIBRATE AND TRANSFER
6. Scaling and the ground metric
If a two-period path is ξ=(arrivals now, arrivals later), raw Euclidean distance treats both coordinates equally. That may be wrong when late demand is expensive because staff cannot be recalled, or early demand triggers a setup decision. A weighted metric can reflect those consequences, but weights should be fixed from operational reasoning rather than tuned until a preferred decision appears.
| Modelling choice | What the same ε permits | Evidence to retain |
|---|---|---|
| Arrivals measured in jobs | Movement in job units | Units and horizon weights |
| Arrivals divided by 100 | Numerically larger relative movement | Scaling transformation |
| Peak-time feature added | Transport can penalise timing shifts | Feature definition and cost link |
| Correlated path coordinates | Whole paths move together | Dependence construction |
CHAPTER 7 OF 12 · CALIBRATE AND TRANSFER
7. Changed-context transfer: two-period readiness
Now let extra capacity ordered after period 1 become available only in period 2. Paths (16,8) and (8,16) have the same total demand and the same distance from (12,12) under some summaries, but they imply different feasible actions. A radius calibrated on total arrivals can therefore protect the wrong uncertainty.
Repeat the analysis with whole paths, preserve peak timing, and evaluate the robust action on the original paths. This extends the work on scenario reduction and tail preservation: both compression and ambiguity must respect the action timeline.
CHAPTER 8 OF 12 · CALIBRATE AND TRANSFER
8. Decision stability across radii
Solve on a grid beginning at ε=0. Record capacity, worst-case objective, empirical objective and held-out cost. The useful object is a stability path: intervals of ε for which the action stays unchanged, followed by the radius at which it jumps.
A jump is not automatically evidence of truth. It may reveal a genuine hedge, or it may expose an arbitrary metric scale. Refit across bootstrap samples and alternative justified ground metrics. If the switch point moves widely, report that instability instead of announcing one robust capacity.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. What the robust guarantee does not say
The guarantee is conditional on the ambiguity set containing relevant distributions and on the loss matching operations. It does not repair a biased sample, invent a missing disruption type or validate an independence assumption. If demand support is physically bounded, encode that support; otherwise an adversary may move small mass to implausibly large values.
Wasserstein robustness is also not the same as a chance constraint or a tail-risk measure. Those may be combined, but each adds its own parameters and interpretation. “Worst case” means worst inside the declared set, not worst imaginable.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Save the sample window, path construction, units, ground metric, support restrictions, radius rule and solver formulation. For each candidate radius, keep the adversarial distribution or transport plan when available. Evaluate the chosen action on held-out observations and on named operational stresses.
Check duality and reformulation assumptions against the loss actually used. A tractable formula for a Lipschitz loss does not transfer automatically to integer recourse, discontinuous service penalties or decision-dependent uncertainty.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Is the largest tested radius the safest choice?
Not necessarily. It may buy robustness to distributions that are implausible while creating avoidable capacity cost. Choose a range from evidence and decision stakes.
Can ε be compared between studies?
Only after matching units, ground metric, dimensionality, scaling and Wasserstein order. A bare radius has little meaning.
Why keep the empirical result?
It shows the price of robustness and reveals whether the robust action responds to uncertainty or merely reproduces the original optimum.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Mohajerin Esfahani and Kuhn develop data-driven distributionally robust optimisation with Wasserstein balls in Mathematical Programming. Fournier and Guillin provide concentration and convergence results for empirical measures. These support the framework; the capacity and transport ledgers above are original teaching examples.
Continue through the Queueing Theory root, scenario reduction and decision stability and joint quantile paths. The Gold Standard of Education explains the transfer check, and How X Works connects this method to other systems.
Top · Contents · Previous · How X Works