Choosing an approach

CriticalTransitions.jl provides several complementary methods that all answer the same underlying question: how, where, and how often does a system undergo a rare transition between metastable states? They differ in the regime where they are valid (noise strength, state-space dimension) and in what they return (a scalar rate, the most probable path, a spatial flux field). This page is the map; each method has its own manual page with the details.

The method menu

Your goalReach forRegime / output
Observe transitions directlyDirect samplingany noise level; brute-force trajectories
Most probable path and exponential rateLarge deviation theoryweak noise ($\varepsilon \to 0$); instanton & quasipotential
Escape rate, mean first-passage time, (quasi-)stationary densityRates, distributions & the generatorfinite noise; spectral observables
Where reactive paths concentrate, transition channels, committorTransition path theoryfinite noise; spatial flux
Transition driven by parameter drift rather than noiseRate-induced transitionsnon-autonomous forcing

Picking by what you need

  • Only a rate or waiting time? Use the spectral observables (mean_first_passage_time, quasi-stationary distribution) on the generator, or Kramers asymptotics for high barriers.
  • The transition path itself? Use large deviation theory (the instanton via MAM / gMAM / sgMAM / shooting).
  • Spatial flux and channels? Use transition path theory.
  • High barrier, weak noise? Direct sampling becomes exponentially expensive; prefer LDT or the spectral observables.
  • Shallow barrier? Just simulate with direct sampling.
  • Parameter drift, not noise? Use rate-induced transitions.

Composing methods

The methods are not mutually exclusive. Transition path theory and the spectral rate observables are both built on the same discretised generator. In large deviation theory, a fast gMAM / sgMAM solve makes a good warm start for the more delicate multiple-shooting boundary-value problem; see Large deviation theory.