Does the method make the finest approximation exact?

No. Standard finite-level MLMC retains the finest approximation's bias relative to the intended exact quantity.

Must every level use the same number of samples?

No. Unequal allocation is central to obtaining efficiency when costs and correction variances differ.

Is the coarse result trusted without checking?

No. Corrections account for systematic differences between levels, and pilot diagnostics assess whether the hierarchy behaves as expected.

AI Glossary / Mathematics and optimization / Advanced

Multilevel Monte Carlo method

Definition

The multilevel Monte Carlo method estimates an expected quantity by combining many inexpensive coarse simulations with fewer carefully paired corrections from finer simulations. It uses a hierarchy of approximations rather than running every sample at the highest resolution. When neighboring levels can be strongly coupled, the corrections may have much lower variance than the full fine-resolution output.

What it means

A simulation often has two distinct errors: random sampling error and approximation bias from its numerical resolution. Running more coarse simulations reduces sampling error but does not remove coarse-resolution bias. Running every sample finely can be costly. Multilevel Monte Carlo estimates a coarse expectation and then adds estimated differences between successive resolutions. The final result targets the expectation at the finest chosen level, retaining that level's approximation bias while potentially spending much less effort on random sampling. It is not simply an average of cheap and expensive answers.

How it works

Let P0, P1 and subsequent quantities denote increasingly accurate approximations, ending at PL. The identity E[PL] = E[P0] + Σ E[Pl − P(l−1)] expresses the fine expectation as a telescoping sum. Estimate each term with its own samples. Within a correction sample, couple the fine and coarse simulations using shared underlying randomness so their difference is stable. Sample sets for different levels can be independent, allowing their estimator variances to add. Pilot runs estimate correction variances and costs; these guide allocation of samples across levels. Choose the finest level with attention to remaining bias as well as the total sampling variance.

Three practical examples

A stochastic engineering model estimates average system response using several time-step sizes, pairing fine and coarse trajectories from the same random disturbances.
A simulation of uncertain material properties combines mesh resolutions when adjacent meshes produce strongly related outputs for the same material sample.
A research team estimates an expectation from a hierarchy of approximate numerical solvers, checking that the coupling and correction variances support a worthwhile multilevel calculation.

Think of it this way

Think of estimating the cost of a large collection of items by making a broad rough assessment and then carefully measuring successive corrections to that assessment. The expensive work focuses on what the rough view missed. The analogy captures correction-based estimation, but mathematical coupling and variance control are essential; casual spot checks do not automatically create a valid multilevel estimator.

Worked example

Consider two levels. Suppose the estimated coarse mean is 10 and the estimated mean of paired fine-minus-coarse differences is 0.4. The multilevel estimate is 10.4. For an illustrative cost calculation, assume a coarse sample has variance 4 and costs 1 unit, while a paired correction has variance 0.04 and costs 10 units. With 1,000 independent coarse samples and 100 independent correction samples, estimator variance is 4/1,000 + 0.04/100 = 0.0044, giving standard error about 0.0663. Total cost is 1,000 × 1 + 100 × 10 = 2,000 units. If a plain fine sample instead has variance 4 and costs 9 units, achieving variance at most 0.0044 needs 910 samples and costs 8,190 units. These stipulated variances and costs illustrate potential savings; they are not universal constants or an optimized allocation. Both approaches still inherit the same finest-level approximation bias.

Illustrative case study

In a fictional modeling project, Brook Dynamics estimates average vibration response under random loading. Its initial pilot finds that paired coarse and fine simulations track one another closely, making the correction variance small. The team allocates many coarse samples and fewer fine corrections, then checks bias by adding a finer level. A different output based on a sharp threshold has less favorable correction behavior, so the team reassesses the method for that output. No real runtime saving is claimed by this scenario.

Common misconceptions and mistakes

Do not run unrelated fine and coarse trajectories and assume their difference will have low variance. Shared randomness must preserve the correct marginal distribution of each approximation. More samples cannot remove an incorrect physical model or unresolved discretization bias. Small average corrections alone are not enough: correction variance also matters. Multilevel Monte Carlo is distinct from multilevel splitting for rare events and from merely parallelizing ordinary Monte Carlo. Its complexity improvements require assumptions about bias, variance decay and simulation cost, rather than applying to every hierarchy automatically.

Business use

Use a pilot study before committing a simulation budget. Record the quantity being estimated, each level's resolution, coupling design, observed correction variance and measured cost. Report sampling uncertainty separately from estimated resolution error. Compare methods at comparable total error, because a cheap but biased answer is not an equivalent result. Check whether rare or discontinuous outcomes undermine the coupling. In repeated studies, preserve random-number handling and model versions so changes in estimated expectations can be distinguished from changes in implementation or sampling noise.

Check your understanding

Doubling every sample count makes sampling error smaller, but the finest numerical grid remains unchanged. Has all error necessarily been halved?

Answer: No. The estimator's variance decreases, its standard error falls only by the square-root factor under ordinary independent sampling, and the finest-grid bias remains.

Related terms

Monte Carlo simulation
Control variates
Root Mean Squared Error
variance reduction
control variate
coupling
discretization bias

Learn next

simulate a small stochastic process using two time-step sizes driven by the same increments. Compare the variance of paired differences with differences from independent paths. Then calculate how changing sample counts affects sampling variance while leaving resolution bias unchanged.

Stochastic process

Sources

Sources: Michael B. Giles, Multilevel Monte Carlo Path Simulation, sections 1–3, author manuscript, https://artowen.su.domains/courses/362/readings/GilesMultilevel.pdf ; author's research overview, https://people.maths.ox.ac.uk/gilesm/mlmc.html . The paper establishes the telescoping estimator and conditional complexity results; the cost example above uses explicitly hypothetical inputs.

Previous: Multilayer perceptron Next: Multilinear principal component analysis