Find your next learning step
A queue-delay exceedance curve must fall as the threshold rises. Yet a finite simulation can produce crossing estimates or band endpoints. Shape constraints can restore coherence, but only a coverage-preserving construction can turn a smoother graph into stronger evidence.
Open the chapter contents · Queueing Theory root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the decision, or read from Chapter 1. The chapters remain open; these grouped routes are a navigable index.
Route 1 · Define the evidence
Route 2 · Work the ledgers
Route 3 · Transfer the method
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE THE EVIDENCE
1. A grid is a decision instrument
A threshold grid samples a curve at operationally meaningful delays. It is not the curve itself. A grid with 30-, 60- and 90-minute points may be sufficient for three service promises, while a scientific description of the complete distribution needs protection between those points as well.
Refinement adds thresholds where a decision boundary, steep change or uncertainty gap matters. Adding points after seeing the weakest-looking location changes the protected family unless the original method covered every threshold from the start.
CHAPTER 2 OF 12 · DEFINE THE EVIDENCE
2. Monotonicity is information, not decoration
For an exceedance curve p(t)=Pr(W>t), larger t cannot have larger probability. Separate Monte Carlo estimates can violate that order because of sampling variation, unequal weights or numerical approximation.
A monotone projection can improve coherence and sometimes precision. The projection must be attached to a theorem or a set operation that preserves simultaneous coverage. Isotonic smoothing of centre lines alone does not automatically repair pointwise intervals.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. A crossing band ledger
Suppose one jointly protected procedure returns the following raw bounds. The upper endpoint at 60 minutes exceeds the one at 45 minutes, although the true curve is non-increasing.
| Threshold | Estimate | Raw lower | Raw upper |
|---|---|---|---|
| 30 min | 8.8% | 7.0% | 11.0% |
| 45 min | 7.2% | 6.1% | 9.5% |
| 60 min | 6.9% | 5.8% | 10.1% |
| 75 min | 4.4% | 3.0% | 6.0% |
The raw band may still cover at every point. The problem is that readers cannot interpret its shape directly, and local decisions may contradict one another.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. Worked order-closure calculation
For a non-increasing curve, a coverage-preserving order closure can use U*(t_j)=min over earlier-or-equal raw upper bounds and L*(t_j)=max over later-or-equal raw lower bounds. Any truly monotone curve inside the raw simultaneous band also lies inside this closed band.
| Threshold | Raw lower | Closed lower | Raw upper | Closed upper |
|---|---|---|---|---|
| 30 min | 7.0% | 7.0% | 11.0% | 11.0% |
| 45 min | 6.1% | 6.1% | 9.5% | 9.5% |
| 60 min | 5.8% | 5.8% | 10.1% | 9.5% |
| 75 min | 3.0% | 3.0% | 6.0% | 6.0% |
This closure uses only logical implications of monotonicity. A least-squares isotonic projection answers a different estimation question and needs its own uncertainty argument.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Refine where the decision can change
A useful refinement rule targets intervals where neighbouring bands straddle an operational cap, where curvature may hide a crossing, or where interpolation error is material. Predeclare the rule or use a band valid over the full continuum.
Do not refine merely because a noisy point looks dramatic. The threshold labels, cap function and stopping rule belong to the statistical design. Grid density cannot compensate for unstable importance weights or missing support in the tail.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Keep time, value and shape guarantees separate
A value-uniform band protects threshold inspection; a time-uniform band protects repeated looks; a shape restriction narrows the feasible curve set. A complete claim must state which combination has actually been proved.
If a fixed-grid familywise method is used, each newly introduced threshold consumes protection. If a time-and-value-uniform cumulative-distribution band is used, adaptive threshold inspection may already be covered inside its stated domain.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Changed-context transfer: call-centre service levels
Transfer the method to a call centre with caps at 20, 60 and 120 seconds. Add a 90-second threshold only when it clarifies a staffing decision or a predeclared refinement rule triggers it.
The transfer check verifies that calls are sampled under the target routing policy, repeated callers and shared shocks are handled, weighted paths retain support, and the service cap is not selected after examining the curve.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Connect the confidence-band branch
The simultaneous-threshold guide defines joint threshold and time protection. The quantile-crossing guide treats coherent forecast vectors and capacity actions.
Use the time-uniform confidence-sequence guide when monitoring can stop after an evidence boundary is crossed.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
Monotonicity does not imply smoothness, convexity or a particular tail family. Imposing extra shape constraints without physical justification can exclude the truth while producing an impressively narrow band.
A dense grid is not automatically more informative. Highly correlated neighbouring points may add visual detail without resolving the decision, while multiplying error if only pointwise intervals are used.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Archive the original grid, refinement rule, curve estimator, weighting scheme, raw simultaneous band, shape constraint, projection or closure operator, cap function, monitoring times and stop rule. Preserve the pre-projection result for auditability.
Check order coherence, effective sample size by threshold, largest weights, band width near every cap and sensitivity to a coarser and finer grid. State whether coverage is fixed-grid, continuum-wide, time-uniform or joint.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Can I project each endpoint separately?
Only when the operation is justified to preserve the claimed coverage. Logical order closure is different from arbitrary curve smoothing.
Does a finer grid always tighten the band?
No. With a fixed global error budget, more grid points can widen componentwise bands. Continuum methods have their own complexity costs.
Why keep the raw band?
It shows what came from sampling evidence and what came from the shape restriction, making the gain and its assumptions visible.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Dümbgen constructs confidence bands for isotonic or convex curves with guaranteed coverage. Guntuboyina and Sen review nonparametric shape-restricted regression. Mineiro and Howard develop time-uniform confidence bands for cumulative distribution functions. The queue calculations above are original teaching examples.
Continue through the Queueing Theory root, the Gold Standard of Education and How X Works.
Top · Contents · Previous · How X Works