Robustness depends on which distributions you allow the model to fear
Distributionally robust optimisation chooses a decision against the worst expected loss inside an explicitly constructed ambiguity set rather than pretending one estimated probability distribution is exact.
Stochastic optimisation usually assumes a probability model for uncertain demand, returns, travel time or failure. With limited data, that model is itself uncertain. A decision that is excellent under the fitted distribution can perform badly when the future distribution differs, even when individual observations look familiar.
Distributionally robust optimisation, or DRO, makes that second layer visible. A Wasserstein ball collects probability distributions within a transport distance of an empirical or nominal distribution. The radius controls caution. The method can provide finite-sample statements under stated assumptions, but no theorem removes misspecification, non-stationarity or the need for external validation.
Chapter contents
- 1. Separate uncertain outcomes from uncertain distributions
- 2. Construct the empirical distribution transparently
- 3. Read Wasserstein distance as transport cost
- 4. Define the ambiguity set and support
- 5. Calibrate the radius rather than treating caution as a mood
- 6. Worked example: robust stocking under demand ambiguity
- 7. Understand worst-case expectation and the inner problem
- 8. Use tractable reformulations only under their assumptions
- 9. Interpret out-of-sample guarantees precisely
- 10. Validate against stochastic, robust and simple baselines
- 11. Audit shift, dependence and protected groups
- 12. Govern deployment, radius changes and human override
Choose a reading route
- Route 1: Frame ambiguity — decision, loss and unknown distribution
- Route 2: Build the ball — empirical centre, Wasserstein distance and radius
- Route 3: Solve a case — worst-case expectation and reformulation
- Route 4: Validate honestly — guarantees, holdouts and sensitivity
- Route 5: Deploy carefully — shift, fairness and review
How to use this guide
Start with the chapters that match your immediate decision, then return to the worked example and audit chapters. Each chapter separates a model from the evidence needed to use it responsibly. A claim is not strengthened merely by being technical; it is strengthened when assumptions, observations, alternatives and limits can be inspected.
| Layer | Question | Minimum record |
|---|---|---|
| Concept | What exactly is being represented? | Definitions and boundary |
| Method | How will the reasoning be carried out? | Steps, inputs and assumptions |
| Transfer | Would it work in a changed setting? | A deliberately altered case |
| Audit | What could prove it wrong? | Evidence, limits and revision trigger |
Chapter 1. Separate uncertain outcomes from uncertain distributions
The idea
Ordinary stochastic optimisation averages loss under one distribution. DRO recognises that the distribution itself is estimated or otherwise ambiguous and optimises against a set of plausible distributions.
A disciplined method
Define decision x, uncertainty ξ, loss ℓ(x,ξ), available observations and the scientific reason the data-generating distribution is uncertain.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which uncertainty belongs in outcomes and which belongs in the distribution model. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A retailer knows demand varies and has only forty historical weeks; both weekly demand and its underlying distribution are uncertain.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Increase sample size while keeping seasonality unstable and distinguish sampling uncertainty from structural change.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: DRO does not automatically model every form of ambiguity, including causal change and adversarial manipulation.
Evidence to retain: decision set, loss, sample, sampling process, exclusions and modelling assumptions
Common misconception: DRO simply adds a safety margin to any optimisation problem.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 2. Construct the empirical distribution transparently
The idea
A common centre is the empirical distribution that places equal mass on observed samples. Its simplicity makes data dependence visible but preserves sampling bias and dependence.
A disciplined method
Document sample selection, weights, time period, missing values and preprocessing; test whether independent and identically distributed assumptions are defensible.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which observations define the nominal centre and with what weights. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Fifty demand observations each receive mass 1/50 after justified cleaning; promotional weeks are retained and labelled rather than silently removed.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Reweight recent observations and show how the ambiguity centre and decision change.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Empirical distributions are discrete and can poorly represent tails, dependence or regime shifts.
Evidence to retain: raw sample, weights, transformations, dependence checks and versioned dataset
Common misconception: The empirical distribution is the true distribution once enough rows are collected.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 3. Read Wasserstein distance as transport cost
The idea
Wasserstein distance measures the minimum cost of moving probability mass from one distribution to another under a chosen ground metric and order. Geometry therefore enters the ambiguity set.
A disciplined method
Scale features, choose the ground metric and norm, justify transport costs and calculate simple one-dimensional moves before using a solver.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which differences in uncertainty space should be considered near or far. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Moving probability mass from demand 100 to 105 costs less than moving it to 180 under absolute distance.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change units from items to thousands without rescaling and observe how the ambiguity set is distorted.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: The metric embeds judgement; inappropriate scaling or mixed data types can make distances meaningless.
Evidence to retain: feature units, scaling rule, metric, Wasserstein order and sensitivity results
Common misconception: Wasserstein distance is a neutral property of the data independent of representation.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 4. Define the ambiguity set and support
The idea
A Wasserstein ball contains distributions whose distance from the centre is no greater than radius ε, often subject to support constraints. The ball is a modelling claim, not a discovered fact.
A disciplined method
State centre, metric, radius, support and moment or structural constraints; check whether implausible distributions enter and plausible ones are excluded.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which distributions the decision is protected against. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A demand model allows non-negative values within a Wasserstein radius but excludes impossible negative demand through support constraints.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Remove the upper support bound and inspect whether the worst case moves mass to unrealistic extremes.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Broad support and unbounded loss can make the robust expectation infinite or overly conservative.
Evidence to retain: complete set definition, support evidence, feasibility tests and example worst-case distributions
Common misconception: Choosing Wasserstein automatically creates a realistic ambiguity set.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 5. Calibrate the radius rather than treating caution as a mood
The idea
The radius controls how far the adversarial distribution may move from the centre. Statistical bounds, cross-validation and task-specific stress tests answer different calibration questions.
A disciplined method
Compare theoretically motivated radii with holdout performance and sensitivity curves; state confidence parameters and select a radius before viewing final test results.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is what evidential claim supports the chosen radius. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A concentration bound yields a conservative radius; validation selects a smaller operational value, and both results are reported rather than blended.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Double the dimension with the same sample size and examine how statistical calibration deteriorates.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Finite-sample bounds can be loose and depend on tail, independence and metric assumptions.
Evidence to retain: candidate radii, derivation, validation split, sensitivity curve and selection rule
Common misconception: A larger radius is always safer and therefore better.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 6. Worked example: robust stocking under demand ambiguity
The idea
A one-product case shows how a decision changes as the ambiguity radius permits more adverse demand distributions.
A disciplined method
Use historical demand, underage and overage costs, construct an empirical centre, solve at ε=0 and several positive radii, and compare holdout cost and stockouts.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which radius gives a decision that balances robustness and cost. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
The empirical solution stocks 110 units; moderate ε raises stock to 118 because the worst case shifts mass toward costly shortages.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Reverse the cost asymmetry so overstock is more expensive and observe that robustness need not always increase quantity.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Illustrative values do not establish a universal optimal radius or direction of change.
Evidence to retain: sample, costs, solver formulation, worst-case distribution, decisions and holdout results
Common misconception: DRO always produces a larger and more conservative decision.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
| Radius | Interpretation | Required check |
|---|---|---|
| 0 | Empirical stochastic model | Baseline fit |
| Small | Local distributional ambiguity | Holdout gain |
| Moderate | Broader stress | Cost of caution |
| Large | Extreme set | Plausibility and finiteness |
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 7. Understand worst-case expectation and the inner problem
The idea
DRO minimises the maximum expected loss over the ambiguity set. The inner adversary selects a distribution, not an arbitrary outcome, and remains constrained by transport and support.
A disciplined method
Write the min-max problem, inspect the worst-case distribution for sample decisions and distinguish distributional robustness from pointwise robust optimisation.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the adversary moves probability mass or selects one extreme value. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
For a fixed stock level, the adversary reallocates limited probability mass toward shortage states while paying transport cost.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Replace the Wasserstein ball with a box uncertainty set over outcomes and compare the different worst cases.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Worst-case distributions can be mathematically valid yet scientifically implausible if the set is poorly designed.
Evidence to retain: primal problem, dual variables, adversarial distribution, support and solver tolerances
Common misconception: DRO assumes the single worst observation occurs with certainty.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 8. Use tractable reformulations only under their assumptions
The idea
Strong duality and problem structure can convert some Wasserstein DRO problems into finite convex, linear or conic programs. Tractability is a theorem with conditions, not a generic feature.
A disciplined method
Match loss regularity, support and Wasserstein order to a published reformulation; document dual derivation, solver class and numerical tolerances; verify on a small primal instance.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which mathematical assumptions make the finite program equivalent. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A piecewise-linear loss under a 1-Wasserstein ball yields a finite convex reformulation; a non-convex recourse model may not.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add integer recourse and identify which equivalence or computational guarantee no longer applies.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Approximations and relaxations can change guarantees even when solver status is optimal.
Evidence to retain: theorem reference, assumptions checklist, algebra, code, solver log and gap
Common misconception: Any Wasserstein DRO model can be solved efficiently with standard linear programming.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 9. Interpret out-of-sample guarantees precisely
The idea
Some data-driven Wasserstein DRO results bound future expected cost with stated confidence when the true distribution lies in the ambiguity set under sampling and tail assumptions.
A disciplined method
Quote the guarantee with probability, sample size, radius rule, loss and assumptions; distinguish an upper confidence bound from realised performance or universal protection.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which random event and probability the theorem actually concerns. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A theorem provides confidence 1−β that the true expected cost is below the robust objective, given independent samples and calibrated ε.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Introduce temporal dependence and explain why the original guarantee cannot be repeated unchanged.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Bounds may be conservative, asymptotic or inapplicable under distribution shift.
Evidence to retain: theorem statement, parameter values, assumption tests and empirical coverage study
Common misconception: An out-of-sample guarantee means no future outcome can be bad.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 10. Validate against stochastic, robust and simple baselines
The idea
Robustness is useful only relative to alternatives and the decision cost. Compare empirical stochastic optimisation, DRO, pointwise robust optimisation and simple policies on untouched or rolling data.
A disciplined method
Pre-register metrics, keep a final test set, evaluate average and tail loss, constraint violations and decision stability, and repeat across time or sites.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether improved tail behaviour justifies average cost and complexity. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
DRO reduces worst decile cost but raises mean cost slightly; the trade-off is shown instead of reported as a single win.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Shift the test period into a new regime and compare which policy fails gracefully.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: One split can flatter any model, and historical tests may not represent future interventions.
Evidence to retain: baselines, splits, seeds, metrics, uncertainty intervals and failure cases
Common misconception: A model with the best robust objective must perform best on unseen data.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 11. Audit shift, dependence and protected groups
The idea
Wasserstein proximity around historical data may not cover policy feedback, causal change or group-specific harm. Aggregate robustness can hide unequal errors or burdens.
A disciplined method
Monitor feature and outcome shift, preserve subgroup performance where lawful, test alternative metrics and support sets, and map worst-case decisions to affected people.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whose distribution and loss the ambiguity set represents. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A scheduling model is robust overall but repeatedly shifts delay to a small worker group because the loss aggregates minutes.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add group-specific service constraints and compare feasibility, cost and robustness.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Sensitive data, small groups and feedback effects complicate auditing and require governance beyond optimisation.
Evidence to retain: drift measures, subgroup metrics, constraint rationale, appeals and review trigger
Common misconception: Distributional robustness guarantees fairness because it considers many distributions.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 12. Govern deployment, radius changes and human override
The idea
A production DRO model is a decision system with versioned data, set design, solvers and consequences. Radius changes and overrides should be reviewable.
A disciplined method
Version every component, log inputs and decisions, define change control, provide an appeal or override route and retire the model when assumptions fail.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which authority may trade performance for robustness and who bears the cost. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A capacity model raises reserves after a radius update; reviewers inspect evidence, operational cost and affected service levels before release.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Transfer the model to a new region and require fresh sampling and calibration rather than copying ε.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Governance cannot make an unsuitable model appropriate; some decisions should remain outside automated optimisation.
Evidence to retain: model card, dataset, ambiguity specification, validation, approvals, overrides and incidents
Common misconception: Once a model has a theorem, later radius changes are routine technical maintenance.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Field workbook
Build a one-dimensional newsvendor example from a small synthetic dataset. Plot decisions and holdout loss over at least five radii, inspect one worst-case distribution, compare two baselines and state exactly which assumptions any guarantee needs.
- Write the decision question in one sentence and name the decision owner.
- List observations separately from interpretations, with dates and provenance.
- Record at least two plausible alternatives before choosing a preferred account.
- Run the changed-context test and note which assumptions fail.
- State the limit, confidence and next review trigger in plain language.
| Layer | Evidence | Failure signal |
|---|---|---|
| Centre | Sample and weights | Biased empirical distribution |
| Geometry | Metric and support | Meaningless transport |
| Radius | Calibration and sensitivity | Caution by intuition |
| Guarantee | Theorem and validation | Claim exceeds assumptions |
Continue through the verified ecosystem
- Bayesian model averaging — model uncertainty through predictive mixtures
- Robust Markov decision processes — policy robustness under transition ambiguity
- Multistage stochastic programming — recourse and scenario trees
- Chance-constrained optimisation — reliability targets and violation budgets
- Gaussian processes — posterior prediction and calibration
Frequently asked questions
What does distributionally robust optimisation protect against?
It protects against the worst expected loss over a stated set of probability distributions, not every imaginable future.
What is a Wasserstein ball?
A set of distributions within a chosen Wasserstein transport distance of a centre distribution.
What does the radius do?
It controls the breadth of distributions considered; larger is usually more conservative but can become unrealistic or costly.
Is DRO the same as robust optimisation?
No. Classical robust optimisation often protects against uncertain outcomes or parameters, while DRO protects against uncertain probability distributions.
Are out-of-sample guarantees unconditional?
No. They depend on the theorem, sampling, tail, loss, support and calibration assumptions.
Does DRO guarantee fairness?
No. Fairness requires explicit objectives, constraints, data, evaluation and governance for affected groups.
Authoritative sources and further reading
- Esfahani and Kuhn: Data-driven DRO using the Wasserstein metric — primary paper on guarantees and tractable reformulations
- Author preprint of data-driven Wasserstein DRO — full primary manuscript
- Wasserstein DRO: Theory and Applications in Machine Learning — author tutorial and theoretical synthesis
- Hanasusanto and Kuhn: Two-stage DRO over Wasserstein balls — conic reformulations and recourse
- Delage and Ye: Distributionally robust optimization under moment uncertainty — foundational ambiguity-set formulation
This guide explains an optimisation framework. Guarantees are conditional on the cited mathematical assumptions and do not replace domain validation, causal reasoning, fairness review or accountable human decisions.