Skip to content

Scenario or Monte Carlo Joint-State Sampling

Simulation method — instantiates Conjunctive Path Assurance

Samples many correlated joint states to estimate how often an entire route conducts at once — the rare-coincidence probability that no single-factor analysis reveals.

Knowing a conjunction can happen does not tell you how often it will. Scenario or Monte Carlo Joint-State Sampling is the quantitative sibling: it draws many random joint states — or full time-histories — from the modelled distributions of each condition, checks how often a complete route conducts, and returns a probability for the whole conjunction. Its defining capability is that it samples the conditions together, with their dependence intact — correlations, common drivers, and overlapping time-windows — so it captures the one thing that makes rare conjunctions dangerous and that a factor-by-factor analysis structurally cannot: the events are not independent, and treating them as independent makes the coincidence look impossible when it is merely rare.

Example

A regional power grid runs near thermal limits during a heat wave, and the hazard "cascading outage" needs several lines to trip within the same short window. Multiplying each line's independent trip probability gives a reassuring one-in-billions figure. But heat is a common driver — it pushes every line toward its limit at once, so the trips are strongly correlated. Monte Carlo sampling draws correlated line loadings, weather, and maintenance-outage states across tens of thousands of simulated days, and estimates a joint-trip probability orders of magnitude higher than the independence assumption implied. It also samples when each protective window is open to estimate the chance the windows overlap. The output is not "possible" or "impossible" but a defensible number — and a pointer to the correlation that drives it.

How it works

Build a distribution for each condition and a dependence structure linking them — correlations, shared drivers, or a copula. Sample the joint state (or a time-series of it), evaluate whether a route conducts under that draw, and repeat until the estimate stabilizes; because the target is a rare event, importance sampling biases the draws toward the dangerous region so the tail can be estimated at all. What distinguishes it from its siblings is that it produces a probability with its dependence made explicit and a ranking of which correlation matters most — a quantitative answer where the analytic and formal methods give only a qualitative one.

Tuning parameters

  • Dependence model — independent versus correlated versus common-driver sampling. This is the highest-leverage dial: assuming independence is precisely what makes a correlated coincidence look impossible.
  • Sample size and rare-event technique — more samples, or importance sampling, sharpen the tail estimate where the hazard lives; too few and the rare-event rate is just noise.
  • Time-window resolution — how finely overlapping windows are modelled; coarse steps can miss a brief simultaneous opening.
  • Distribution fidelity — measured shapes versus assumed-normal; a fat tail modelled as thin hides the dangerous region.
  • Scenario versus random — hand-built stress scenarios probe named worst cases; random sampling finds the coincidences no one imagined.

When it helps, and when it misleads

Its strength is that it quantifies the simultaneous conjunction and, above all, represents the correlation and timing overlap that make an independence-based "one in a billion" wrong by orders of magnitude — turning "surely that can't happen at once" into a number with its assumptions on the table.

Its failure mode is garbage-in with a confident face: assume independence where conditions move together, or a thin tail where the real one is fat, and it returns a precise, polished, wrong probability whose 50,000-run histogram hides the guesswork in its priors.[1] Rare events also demand enormous samples or careful importance sampling to estimate at all. The classic misuse is tuning the priors until the risk clears the threshold — sampling to the answer you wanted. The discipline is to feed measured dependence, validate the estimate against history where it exists, and carry its uncertainty forward rather than reporting the point estimate as fact.

How it implements the components

  • probabilistic_dependence_profile — sampling the conditions jointly with a modelled correlation or common-driver structure is this profile: its central input and its most consequential output.
  • temporal_overlap_model — sampling when each condition's window is open estimates the probability that the required windows coincide, the timing half of the conjunction.

It does not enumerate which combinations to worry about — that ranked list is Minimal Cut-Set Enumeration — nor prove a combination impossible, which is Model Checking and Reachability Analysis; it prices the conjunctions others identify.

Notes

Reach for sampling only when likelihood matters and the conditions are genuinely dependent. If they really are independent, a closed-form product of probabilities gives the same answer far more cheaply, and the simulation's extra machinery merely adds places for a wrong prior to hide. Its worth is decided by the dependence structure it is fed, not by the number of iterations it runs.

References

[1] Multiplying the individual probabilities of events assumes they are independent; when a common driver correlates them, the true joint probability is far higher than the product — the reason a "perfect storm" is more frequent than an independence calculation predicts. Modelling that dependence is the whole reason to sample rather than multiply. ↩