Find your next learning step
One aggressive tilt may find one rare delay and miss another. A proposal mixture can span several congestion mechanisms while a defensive baseline component protects support and limits how extreme a return weight can become.
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 · Frame the problem
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 · FRAME THE PROBLEM
1. Why one proposal rarely covers every queue tail
Rare-delay simulation changes the distribution of arrivals, services or disruptions so consequential paths occur often enough to study. A proposal tuned to a 20-minute wait can be efficient near that threshold yet poor for a 60-minute wait driven by a different mechanism. One tilt may create many moderate bursts; another may reveal long service stalls.
A mixture proposal samples from several designs. Its denominator is the density of the full mixture, not merely the component that happened to generate the path. The aim is not to average several convenient answers. It is to create one auditable sampling law whose coverage can be checked across the decision thresholds that matter.
CHAPTER 2 OF 12 · FRAME THE PROBLEM
2. The mixture denominator and the defensive component
Let the baseline path density be f and let proposal components be q₁,…,qJ with mixture shares α₁,…,αJ. The sampled density is qmix(x)=Σαjqj(x), so the return weight is f(x)/qmix(x). Using f(x)/qj(x) for a path drawn from component j ignores the other components and generally gives the wrong estimator.
A defensive mixture gives the baseline itself a positive share α₀. Then qmix(x) is at least α₀f(x), so any ordinary importance ratio is at most 1/α₀ wherever f is defined. This protects support and weight magnitude, but it spends some runs on baseline paths that may contain few severe delays.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. A four-component design ledger
Suppose 10,000 paths are divided among baseline sampling, a moderate-arrival tilt, a severe-arrival tilt and a service-stall tilt. The allocation is chosen before the main run and every component density can be evaluated for every path.
| Component | Mixture share | Paths | Primary coverage role |
|---|---|---|---|
| Baseline f | 0.10 | 1,000 | Support defence and ordinary traffic |
| Moderate-arrival tilt | 0.35 | 3,500 | 15–30 minute waits |
| Severe-arrival tilt | 0.35 | 3,500 | 30–60 minute waits |
| Service-stall tilt | 0.20 | 2,000 | Long waits from slow recovery |
Because the baseline share is 0.10, the mixture weight is bounded above by 10 under the stated density convention. That bound does not guarantee low variance: weights can still concentrate near ten, and an outcome that no non-baseline component reaches may remain expensive to estimate.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. A threshold-by-threshold evidence ledger
Audit each delay threshold rather than giving the whole mixture one quality label. Consider a run whose estimates and diagnostics look as follows. Effective scenario count is a weight-concentration summary for the relevant tail contribution, not a literal number of independent customers.
| Delay threshold | Tail hits | Largest normalised contribution | Effective-count signal |
|---|---|---|---|
| 15 minutes | 1,420 | 1.2% | Strong |
| 30 minutes | 410 | 4.8% | Usable |
| 45 minutes | 96 | 13.5% | Fragile |
| 60 minutes | 18 | 31.0% | Unreliable without redesign |
The proposal covers the first two thresholds well, but the 60-minute estimate is carried by too few influential paths. Reporting one overall Monte Carlo error would hide the point at which the evidence stops supporting action.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Choosing mixture shares without chasing the answer
A pilot can reveal which components reach which thresholds, but selecting shares from the same outcomes later reported creates adaptation that must be accounted for. Use a separated pilot or a valid adaptive importance-sampling scheme, freeze the final mixture and retain the pilot decision record.
Allocate by decision need as well as statistical efficiency. If a safety rule is triggered at 60 minutes, a mixture that estimates only the 20-minute tail precisely is poorly aligned. A small protective component for a second rare mechanism may be worth more than a marginal improvement at the most common threshold.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Balance weighting across several proposals
When paths are drawn from multiple proposals, the balance heuristic uses the mixture density in the denominator and naturally gives less influence where many components provide coverage. It also avoids treating a path as though only its generating component could have produced it.
Evaluate all component log densities with stable log-sum-exp arithmetic. A bookkeeping shortcut that records only the selected component density prevents correct mixture weighting later. Store the component identifier, every evaluable log density, the mixture shares and the baseline log density for each path.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Changed-context transfer: an emergency department
Transfer the design to an emergency department. Components may tilt arrival surges, acuity mix and temporary service reductions. The baseline component protects ordinary combinations, while specialised components expose distinct routes to clinically important waits.
The mathematics transfers; the decision context does not. Patient safety thresholds, diversion rules, staffing constraints and dependence between acuity and service time need domain review. A mixture that covers numerical tails is not proof that the operational model represents care delivery.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Connecting mixtures to existing queue diagnostics
The likelihood-ratio diagnostics guide establishes weight direction, effective scenario count and tilt stability. The Pareto-tail guide examines the most influential ratios. A proposal mixture acts earlier: it changes how the paths are generated.
Use scenario reduction and tail preservation only after distinguishing sampling design from scenario compression. Both can improve computation, but they solve different problems and require different audits.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and misconceptions
A defensive component does not automatically make an estimate precise. The upper weight bound can be large when the baseline share is tiny, and the event indicator can still be rare under every component. Components with support holes, reversed ratios or misnormalised densities remain invalid.
More components are not always better. Redundant components divide simulation effort and increase density-evaluation cost. A mixture tuned after inspecting final results can exaggerate success. Coverage across observed thresholds does not establish coverage for a still more extreme, structurally different event.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Archive the baseline model, component definitions, mixture shares, selection rule, pilot separation, random seeds and per-path component log densities. Verify that shares sum to one, the baseline share is positive when a defensive guarantee is claimed and the mixture denominator is used for every path.
For each threshold, report hits by component, raw and smoothed weights, largest contributions, effective count, Pareto-tail diagnostics, replicate uncertainty and decision stability. Compare with baseline simulation in a less-rare regime and rerun with plausible alternative shares.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Does a baseline component remove all support problems?
It protects support only when the baseline density can be sampled and evaluated correctly and receives positive mixture mass. Model misspecification remains.
Should every threshold get its own proposal?
Not necessarily. One component may cover several thresholds. Add a component when it represents a genuinely distinct influential region and improves the declared evidence target.
Can mixture shares be adapted during the run?
Yes, under an adaptive method with valid weighting and preserved history. Informal share changes without recording the evolving proposal can invalidate the estimator.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Hesterberg develops weighted-average importance sampling and defensive mixture distributions. Owen and Zhou present safe and effective importance sampling with mixture ideas. Vehtari and colleagues provide Pareto-smoothed importance sampling diagnostics. The ledgers above are original teaching examples.
Continue through the Queueing Theory root, Pareto-tail diagnostics and confidence bands and decision margins. The Gold Standard of Education and How X Works provide the wider route.
Top · Contents · Previous · How X Works