Multilevel Monte Carlo Simulation
Also known as: MLMC, multilevel MC, multi-level Monte Carlo, MLMC simulation
Multilevel Monte Carlo (MLMC) is a variance-reduction technique that estimates expectations by combining simulations run at multiple levels of numerical resolution. Coarse, cheap simulations capture most of the signal; fine, expensive simulations correct only the remaining small difference — dramatically reducing total computational cost compared with standard Monte Carlo at the finest level alone.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use MLMC when (1) accurate simulation of a stochastic model requires fine numerical resolution that makes standard Monte Carlo prohibitively expensive; (2) the model admits a natural hierarchy of approximations (finer time steps, finer spatial grids, or higher-fidelity physics) such that consecutive-level outputs are correlated; (3) the target quantity is an expectation (mean, variance, quantile) of a scalar or vector functional. MLMC is particularly powerful for stochastic differential equations, elliptic PDEs with random coefficients, and computational finance. Do not use it when no meaningful hierarchy of approximations exists, when level costs cannot be controlled, or when the variance of level corrections does not decrease as the level refinement increases.
Strengths & limitations
- Reduces computational cost by orders of magnitude relative to standard Monte Carlo at a fixed target accuracy.
- Straightforward to implement once a simulator exists: only paired samples at two consecutive resolutions are needed.
- Provably optimal sample allocation given variance and cost estimates at each level.
- Applicable across diverse domains: SDEs, PDEs with random inputs, agent-based models, and Bayesian inverse problems.
- Naturally parallelisable — samples at each level are independent and can be distributed across processors.
- Extends cleanly to multilevel MCMC and multilevel sequential Monte Carlo for Bayesian inference problems.
- Requires that the model supports a hierarchy of numerical resolutions with increasing accuracy and increasing cost.
- The variance of corrections must decay sufficiently fast across levels; if it does not, the theoretical cost savings vanish.
- Estimating the variance and cost parameters needed for optimal allocation requires a pilot run.
- Implementation complexity grows when the coupling between levels is difficult (e.g. path-dependent payoffs with discontinuities).
- Not directly applicable to models where only a single level of simulation is possible or cost differences across resolutions are negligible.
Frequently asked
How does MLMC differ from standard Monte Carlo?
Standard Monte Carlo runs all N samples at the single target resolution, with cost O(N * C_L). MLMC decomposes the estimation into a telescoping sum of corrections across levels, running most samples cheaply at coarse levels and only a few at the fine level. This can reduce total cost from O(epsilon^{-3}) to O(epsilon^{-2}) for typical SDEs.
What does 'coupled samples' mean in MLMC?
At each level l, a pair of simulations — one at resolution l and one at resolution l-1 — is driven by the same underlying random numbers. This positive correlation ensures the correction P_l - P_{l-1} has much lower variance than either P_l or P_{l-1} alone, which is the source of MLMC's efficiency.
How many levels L should I use?
L is chosen so that the bias from the coarsest approximation is negligible relative to the statistical error target. In practice, start with L such that the weak error at the coarsest level is comparable to epsilon, then add levels until the finest-level correction is negligible. Giles (2015) gives practical guidance on bias and variance estimation.
Can MLMC be used for Bayesian inference?
Yes. Multilevel MCMC and multilevel sequential Monte Carlo extend MLMC ideas to posterior sampling. In these methods, MCMC or SMC chains are run at multiple resolution levels and corrections between levels are estimated, achieving similar cost reductions for Bayesian inverse problems with expensive forward models.
Does MLMC work for discrete-event or agent-based simulations?
It can, but defining a natural coupling between resolution levels is harder. Researchers have used MLMC with agent-based epidemiological and ecological models by defining coarse levels via population aggregation. The key requirement — that variance of corrections decreases with level — must be verified empirically.
Sources
- Giles, M. B. (2008). Multilevel Monte Carlo path simulation. Operations Research, 56(3), 607–617. DOI: 10.1287/opre.1070.0496 ↗
- Giles, M. B. (2015). Multilevel Monte Carlo methods. Acta Numerica, 24, 259–328. DOI: 10.1017/s096249291500001x ↗
How to cite this page
ScholarGate. (2026, June 3). Multilevel Monte Carlo Simulation. ScholarGate. https://scholargate.app/en/bayesian/multilevel-monte-carlo-simulation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Markov Chain Monte CarloSimulation↔ compare
- MONTE-CARLO-SIMULATIONDecision-making↔ compare
- Particle FilterBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare