How Multilevel Monte Carlo Algorithms Cut Derivative-Pricing Cost: Telescoping Estimators, Coupled Fine/Coarse Paths, Bias–Variance Allocation, Complexity and Convergence Failure

Reader question: If a derivative-pricing model becomes more accurate only when we simulate very fine time steps, do we really need to run every Monte Carlo path at the finest and most expensive resolution?

Multilevel Monte Carlo (MLMC) says no. Instead of estimating the finest-resolution expectation directly, it rewrites that expectation as a telescoping sum of corrections between increasingly fine discretisations. Cheap coarse paths estimate the large, easy part of the value. Progressively fewer expensive fine paths estimate only the small corrections that coarse simulation misses.

This article owns one precise computational job: cross-resolution Monte Carlo complexity reduction for discretised stochastic models. Its inputs are a hierarchy of numerical approximations, a coupling rule between neighbouring levels and a target accuracy. Its outputs are a multilevel price estimate, bias estimate, variance allocation, level diagnostics and computational-complexity evidence.

It does not own ordinary Monte Carlo pricing, quasi-Monte Carlo point construction, importance sampling, stochastic-model calibration or a specific derivative payoff. Those are neighbouring jobs already represented in the Bukit Timah Tutor finance-and-banking-algorithms estate.

This is public mathematical and computational education. It is not financial advice, a trading recommendation or a guarantee that an efficient numerical method makes the underlying financial model correct.

1. The ordinary Monte Carlo problem

Suppose the exact quantity of interest is:

P = E[P(X)],

where X is a stochastic process and P(·) is a discounted payoff functional.

We cannot simulate X exactly, so we discretise time. Let P_L be the payoff computed on a fine grid with step size:

h_L = T / 2^L.

A standard Monte Carlo estimator is:

P̂_MC = (1/N) Σ P_L⁽ⁱ⁾.

Its total error contains two different pieces:

• discretisation bias: E[P_L] − E[P];
• sampling error: P̂_MC − E[P_L].

Reducing bias requires a fine grid. Reducing sampling error requires many paths. Running many paths on a very fine grid is exactly what makes ordinary Monte Carlo expensive.

2. The telescoping identity

MLMC introduces a hierarchy:

P₀, P₁, …, P_L

from coarse to fine.

The key identity is:

E[P_L] = E[P₀] + Σ_ℓ=1^L E[P_ℓ − P_ℓ−1].

Nothing has been approximated yet. This is exact algebra: all intermediate expectations cancel.

Instead of estimating E[P_L] directly, MLMC estimates every term separately.

3. Why the identity can save work

The coarse payoff P₀ can differ greatly from zero, so many samples may be needed to estimate its expectation accurately. Fortunately, coarse samples are cheap.

At high levels, P_ℓ and P_ℓ−1 should be very similar if they are generated from the same underlying random path. Therefore their difference:

Y_ℓ = P_ℓ − P_ℓ−1

can have small variance even though each payoff individually remains volatile.

That means only a modest number of expensive fine-level samples may be required.

4. Coupling is the centre of the algorithm

Consider an SDE driven by Brownian increments. On fine level ℓ, two consecutive increments might be:

ΔW_2j, ΔW_2j+1.

The corresponding coarse increment should be:

ΔW_j^coarse = ΔW_2j + ΔW_2j+1.

Fine and coarse paths therefore share exactly the same Brownian movement over each coarse interval.

This produces strong positive correlation between P_ℓ and P_ℓ−1, reducing:

Var(P_ℓ − P_ℓ−1).

If fine and coarse paths were simulated independently, the correction variance would generally remain large and the multilevel advantage would largely disappear.

5. The MLMC estimator

Let N_ℓ be the number of samples on level ℓ.

For level 0:

Y₀ = P₀.

For ℓ ≥ 1:

Y_ℓ = P_ℓ − P_ℓ−1.

The estimator is:

P̂_MLMC = Σ_ℓ=0^L (1/N_ℓ) Σ_i=1^Nℓ Y_ℓ⁽ⁱ⁾.

Each level can use an independent set of paired fine/coarse samples. The dependence that matters is within each pair, not across levels.

6. Bias and variance are controlled separately

The finest level L controls discretisation bias:

|E[P_L − P]|.

The sample counts N_ℓ control sampling variance:

Var(P̂_MLMC) = Σ V_ℓ/N_ℓ,

where:

V_ℓ = Var(Y_ℓ).

A target mean-square error ε² is often split approximately into:

• bias² ≤ ε²/2;
• sampling variance ≤ ε²/2.

The split is a design choice, but the conceptual separation is fundamental.

7. Optimal sample allocation

Let C_ℓ be the expected computational cost of one level-ℓ correction sample.

To minimise total cost:

Cost = Σ N_ℓC_ℓ

subject to a target sampling variance, the continuous optimisation gives the familiar proportionality:

N_ℓ ∝ √(V_ℓ/C_ℓ).

More precisely, after normalising to the variance target, levels receive many samples when variance is high and cost is low, and few samples when variance is low and cost is high.

This is why MLMC should not use the same number of paths at every level.

8. Complexity rates

MLMC theory is often described through three rates:

• α: weak-error/bias rate, |E[P_ℓ−P]| = O(h_ℓ^α);
• β: variance-decay rate, Var(Y_ℓ) = O(h_ℓ^β);
• γ: cost-growth rate, C_ℓ = O(h_ℓ^−γ).

The relationship between β and γ is especially important.

If correction variance falls faster than per-sample cost rises, MLMC can approach the ideal Monte Carlo complexity:

O(ε⁻²).

In Giles’ original Euler/Lipschitz setting, MLMC reduced a representative standard-Monte-Carlo cost of order O(ε⁻³) to O(ε⁻²(log ε)²).

The exact rates depend on the discretisation, payoff and coupling. They must be measured, not memorised as universal constants.

9. Pilot simulation

A practical implementation usually begins with a small number of samples on each existing level.

For each level, estimate:

• mean correction |E[Y_ℓ]|;
• correction variance V_ℓ;
• cost C_ℓ;
• observed decay rates.

These pilot estimates determine both whether another finer level is needed and how future samples should be allocated.

10. A minimal adaptive MLMC algorithm

1. Choose initial levels 0,…,L.
2. Run pilot samples on every level.
3. Estimate correction means, variances and costs.
4. Estimate remaining discretisation bias.
5. If bias is too large, add a finer level.
6. Compute approximately optimal N_ℓ for the variance target.
7. Generate missing samples at each level.
8. Re-estimate level statistics.
9. Repeat until both bias and sampling-error criteria pass.
10. Return price, error estimate, level table and cost report.

11. Inputs and outputs

Inputs can include:

• stochastic process and calibrated parameters;
• payoff function;
• discounting rules;
• base timestep;
• refinement ratio;
• path discretisation (Euler, Milstein, exact bridge, etc.);
• fine/coarse coupling method;
• target RMS error ε;
• pilot sample count;
• maximum level;
• random-number seed policy;
• variance-reduction controls.

Outputs can include:

• MLMC price estimate;
• estimated bias;
• estimated sampling standard error;
• level means E[Y_ℓ];
• level variances V_ℓ;
• level costs C_ℓ;
• allocated sample counts N_ℓ;
• observed convergence rates;
• total runtime/cost;
• comparison with standard Monte Carlo.

12. Evidence polarity

Evidence for confidence includes:

• level-correction means decay smoothly toward zero;
• correction variances decay materially with refinement;
• coupled fine/coarse path differences shrink as expected;
• bias estimates agree with observed level increments;
• estimated cost meets the requested error target;
• repeated runs achieve errors consistent with reported confidence intervals;
• prices agree with analytical/tree/PDE benchmarks where available;
• MLMC cost is lower than standard Monte Carlo at equal accuracy.

Evidence against confidence includes:

• correction variance stops decaying;
• level means oscillate or fail to approach zero;
• results depend strongly on pilot-sample randomness;
• the finest-level correction remains a large fraction of the total price;
• coupling errors create discontinuities;
• claimed error tolerance is not achieved in repeated benchmark runs;
• fine-level sample counts explode unexpectedly.

13. Counterexample: independent fine and coarse paths

Suppose P_ℓ and P_ℓ−1 are generated independently.

Then:

Var(P_ℓ − P_ℓ−1) = Var(P_ℓ) + Var(P_ℓ−1)

when they are independent.

The correction variance may remain roughly as large as the original payoff variance.

Falsifier: compare correction variance under independent and coupled paths. If coupling does not reduce variance materially, the implementation or payoff structure deserves investigation.

14. Counterexample: discontinuous payoff

Digital and barrier payoffs can change abruptly when a fine path crosses a threshold but a coarse path does not.

Even when paths are close, payoff differences can remain large near the discontinuity. This slows variance decay.

Falsifier: plot Var(Y_ℓ) against h_ℓ. If β is weak, use a better coupling, conditional expectation, Brownian-bridge correction or another specialised construction.

See Brownian-bridge barrier algorithms for one related crossing problem.

15. Counterexample: bias is mistaken for sampling error

A simulation can produce an extremely small standard error while still using a grid that is too coarse.

For example:

estimate = 5.000 ± 0.002

may look precise even if the infinite-resolution model price is 5.080.

Falsifier: inspect the level-correction means and add finer levels. Monte Carlo confidence intervals do not automatically include discretisation bias.

16. Counterexample: strong convergence fails near model boundaries

Some stochastic models are difficult to discretise accurately. Square-root variance processes, jumps, reflecting boundaries or path constraints can create negative states or unstable coarse/fine differences under naive schemes.

Falsifier: compare alternative numerical schemes and check path invariants. MLMC amplifies a bad discretisation hierarchy rather than repairing it.

17. Milstein can improve variance decay

When a Milstein discretisation is practical, its stronger pathwise convergence can make fine and coarse paths much closer than under Euler.

This can increase β and improve complexity.

But the Milstein method can require Lévy-area terms in multidimensional systems. Special antithetic MLMC constructions can recover favourable rates without explicitly simulating all Lévy areas in some settings.

The broader lesson is that MLMC performance depends on the strength of the coupling, not merely on having multiple grids.

18. MLMC and quasi-Monte Carlo are different dimensions of improvement

MLMC changes resolution allocation. Quasi-Monte Carlo changes sample-point structure.

They can be combined into multilevel quasi-Monte Carlo, but neither subsumes the other.

See Sobol quasi-Monte Carlo algorithms for the low-discrepancy side of the problem.

19. MLMC and importance sampling are also complementary

Importance sampling changes the probability measure or sampling distribution to spend more effort on important rare events.

MLMC decides how much effort belongs at each numerical resolution.

In rare-event problems, carefully designed multilevel importance sampling can use both ideas, but likelihood-ratio consistency across levels becomes another weak link.

See importance-sampling algorithms.

20. Why MLMC is especially useful for Greeks and nested calculations

Derivative pricing is only one use. Similar multilevel decompositions can estimate:

• sensitivities;
• expected exposure;
• risk measures;
• nested conditional expectations;
• stochastic-PDE quantities.

But every new quantity requires its own coupling analysis. A coupling that works well for price may fail for a discontinuous Greek estimator.

21. Weak links

• fine/coarse random streams are not actually coupled;
• coarse path construction does not match the fine path over common intervals;
• payoff discontinuity destroys variance decay;
• bias estimation uses too few levels;
• pilot variance estimates are unstable;
• random-number reuse creates unintended cross-level dependence;
• cost model ignores memory/vectorisation/GPU effects;
• fine-level discretisation itself is biased or unstable;
• stopping criteria confuse statistical and discretisation errors.

22. What would falsify confidence?

Confidence should be withdrawn if correction means do not approach zero; if correction variances fail to decay under refinement; if empirical errors exceed the claimed tolerance in benchmark repetitions; if coupled and independent corrections have similar variance; if adding fine levels changes the price materially beyond the bias estimate; or if a standard Monte Carlo benchmark achieves the same accuracy at similar or lower cost.

23. Verification and update triggers

Preserve model version, discretisation scheme, level hierarchy, random-number policy, coupling code, pilot estimates, sample allocation, observed α/β/γ rates, bias estimate and benchmark results.

Revalidate when:

• the payoff changes;
• the stochastic model changes;
• the path discretisation changes;
• barrier/jump handling changes;
• random-number libraries change;
• hardware/vectorisation changes the cost profile;
• observed variance-decay rates deteriorate;
• the requested accuracy moves into a much finer regime.

24. Connections to the surrounding knowledge estate

Monte Carlo pricing algorithms own the baseline path-simulation estimator.

Sobol QMC owns low-discrepancy sample design.

Andersen QE shows why the discretisation itself can be the difficult object in stochastic-volatility simulation.

Andersen–Broadie primal–dual algorithms illustrate nested simulation, where multilevel ideas can become especially valuable.

25. Primary and high-quality references

• Michael B. Giles, Multilevel Monte Carlo Path Simulation, Operations Research, 2008.
• Michael B. Giles, Multilevel Monte Carlo Methods, Oxford Mathematical Institute research page and publication archive.
• Michael B. Giles, Multilevel Monte Carlo Methods, Acta Numerica, 2015 review.
• Oxford University Research Archive, Multilevel Monte Carlo Path Simulation, institutional record for the original paper.
• Federal Reserve, Supervisory Guidance on Model Risk Management, for independent validation and ongoing monitoring principles.

Educational boundary: MLMC reduces numerical cost by exploiting resolution structure. It does not remove model risk, market-data risk or the need to verify the finest-level discretisation.