Find your next learning step
A tilted simulation can produce many paths while only a few carry the baseline evidence. Likelihood-ratio diagnostics reveal whether the apparent sample size is real or decorative.
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 a successful tilt can still be fragile
Exponential tilting deliberately samples queue paths from a distribution Q that visits rare delays more often than the baseline distribution P. A baseline expectation is recovered with the likelihood ratio w=dP/dQ. The estimator can be unbiased while remaining practically unstable: most simulated paths may receive almost no weight, and a few paths may control the answer.
That distinction matters after choosing an entropy budget or risk-sensitive tilt. A rare-delay estimator that reports only its mean hides whether the result came from broad evidence or one fortunate trajectory. The first diagnostic question is therefore not “How many paths ran?” but “How many weighted paths actually contributed?”
CHAPTER 2 OF 12 · FRAME THE PROBLEM
2. Weights, normalisation and the target estimand
For paths X_i drawn from Q, estimate a baseline mean E_P[g(X)] by the average of w_i g(X_i), with w_i=p(X_i)/q(X_i). A self-normalised estimate divides the weighted sum by the sum of weights. The two estimators have different finite-sample behaviour; the self-normalised form is convenient when the normalising constant is unknown but is generally biased at finite sample size.
Keep the direction visible. If Q oversamples severe delay paths, dP/dQ usually downweights them when returning to P. Reversing the ratio estimates the wrong measure. Record log weights before exponentiating, subtract a common maximum for numerical stability, and state whether the target is a mean delay, tail probability or policy comparison.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. A six-path likelihood-ratio ledger
Suppose a tilted queue run produces six complete paths. Their baseline-return weights are 1.80, 1.20, 0.70, 0.40, 0.25 and 0.15; their delay losses are 2, 4, 8, 12, 18 and 25. The weights sum to 4.50, while the weighted loss sum is 27.05. The self-normalised baseline loss estimate is therefore about 6.01.
| Path | Delay loss | Likelihood ratio dP/dQ | Weighted loss |
|---|---|---|---|
| 1 | 2 | 1.80 | 3.60 |
| 2 | 4 | 1.20 | 4.80 |
| 3 | 8 | 0.70 | 5.60 |
| 4 | 12 | 0.40 | 4.80 |
| 5 | 18 | 0.25 | 4.50 |
| 6 | 25 | 0.15 | 3.75 |
The largest delay is scientifically important yet does not dominate this baseline mean, because Q was built to see such paths more often. For estimating a rare-tail probability under P, the same sample may behave differently. Diagnostics belong to an estimand, not merely to a file of weights.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. Effective scenario count from weight concentration
A common importance-sampling diagnostic is ESS=(sum w_i)^2/sum w_i^2. Here the sum of squared weights is 5.415, so ESS is about 3.74 even though six paths were simulated. Normalised weights also show that the largest path contribution is 0.40 and the smallest about 0.033. The raw count of six overstates the diversity supporting the estimate.
Call this an effective scenario count when explaining it operationally. It is a concentration summary, not a literal claim that the sample equals 3.74 independent baseline paths. Autocorrelation, target-specific variance and an unstable normalising constant can make the true information smaller.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. A tilt-stability ladder
Run the same decision under mild, medium and strong tilts, with independent seeds. A useful ladder reports effective scenario count, maximum normalised weight, tail hits and the action chosen. If the estimated loss moves modestly while effective scenario count collapses, the extra decimal places are not evidence of stability.
| Tilt | Raw paths | Effective scenario count | Largest normalised weight | Estimated baseline loss |
|---|---|---|---|---|
| Mild | 2,000 | 1,120 | 0.7% | 5.8 |
| Medium | 2,000 | 410 | 2.6% | 6.0 |
| Strong | 2,000 | 34 | 18.4% | 8.9 |
A strong tilt can still be the right tool for a specific rare event. The warning is that one proposal should not be expected to estimate every quantity well. Separate proposals, mixtures or adaptive tilts may be needed for different thresholds and policies.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. From diagnostic numbers to a capacity decision
For each capacity action, estimate both nominal performance and the stressed quantity that motivated the tilt. Then repeat across tilt parameters and seeds. A robust recommendation should not switch because one run contains a single enormous likelihood ratio. If it does, increase simulation, redesign the proposal or admit that the decision margin is unresolved.
Pair statistical stability with operational sensitivity. A stable mean delay can coexist with an unstable ninety-ninth percentile, and two actions can be well estimated yet nearly tied. Report Monte Carlo error beside the real-world decision gap rather than treating a large effective scenario count as a pass certificate.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Changed-context transfer: emergency arrivals
Transfer the method from a cloud-service queue to an emergency department. The rare event is now a long surge of high-acuity arrivals; the action is whether to call an extra team. Preserve complete paths, the proposal probability and the baseline probability. Recompute likelihood ratios for the clinical arrival model rather than carrying weights across domains.
The changed context adds triage classes, time-varying service and ethical constraints. The diagnostic logic transfers: check weight concentration, tail coverage and action stability. The numerical tilt, loss function and acceptable uncertainty do not. This is transfer through structure, not through copied parameters.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Relation to entropy budgets and adversarial paths
The preceding relative-entropy budget article uses exponential tilting to identify costly delay paths. This article asks whether simulations under that tilt support a stable numerical claim. The adversarial distribution for robust control and the proposal distribution for importance sampling can coincide algebraically, yet they answer different questions.
Use entropic risk and rare-delay penalties for the policy objective, and use likelihood-ratio diagnostics for the numerical evidence. Do not call a worst penalised distribution a forecast merely because it generated the simulation paths.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
A high effective scenario count does not prove unbiased code, correct likelihood ratios or adequate rare-event coverage. A low count is a warning, not an automatic rejection: a targeted estimator can still be useful if its variance and error are quantified. Weight clipping improves numerical behaviour but changes the estimand unless the induced bias is handled.
The familiar ESS formula is a heuristic approximation with assumptions. It does not account automatically for dependence among paths, multiple proposals or the particular function g. Always add replicate variance, seed stability, log-weight plots and a direct check against a simpler regime where brute-force simulation is possible.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Archive the baseline and proposal densities, random seeds, horizon, event definition, log likelihood ratios and the exact estimator. Confirm that likelihood ratios average toward one under Q when theory permits. Compare weighted and unweighted summaries so the direction of reweighting is visible.
Report raw path count, effective scenario count, maximum normalised weight, weight quantiles, tail hits, Monte Carlo standard error and policy switching points. Repeat with a defensive mixture containing some baseline sampling. If the decision changes materially, enlarge the evidence rather than averaging incompatible conclusions.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Is effective scenario count a confidence interval?
No. It summarises weight concentration. A confidence interval or replicate error analysis still has to match the estimator and dependence structure.
Should every severe path have a large likelihood ratio?
No. A severe path deliberately made common under Q can have a small dP/dQ ratio. Severity and baseline-return weight are different quantities.
Can I choose the tilt after seeing the answer?
Adaptive selection is possible, but the adaptation and its uncertainty must be part of the method. Unreported tuning invites optimistic stability.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Chen and Choe study importance sampling and diagnostics in stochastic simulation. Elvira, Martino and Robert explain why the usual effective-sample-size approximation has important limitations. Glynn and Iglehart discuss alternative measures and likelihood-ratio behaviour for Markov-chain importance sampling. The six-path and stability ledgers above are original teaching examples.
Continue through the Queueing Theory root, relative-entropy budgets and adversarial delay paths, scenario reduction and tail preservation and Wasserstein ambiguity sets. The Gold Standard of Education and How X Works provide the wider route.
Top · Contents · Previous · How X Works