Nested Monte Carlo (NMC) Methods

• Nested Monte Carlo (NMC) is a stochastic simulation framework for evaluating complex, layered expectations where inner and outer simulations interact nonlinearly.
• Advanced variants like multilevel strategies, unbiased randomized estimators, and regression control variates reduce bias and improve convergence rates.
• NMC methods are applied in fields such as risk assessment, quantum circuit design, Bayesian inference, and computational physics, balancing computational cost with precision.

A nested Monte Carlo (NMC) approach is a class of stochastic simulation techniques designed to estimate quantities expressed as complex, especially compositional or iterated, expectations. These methods are essential when the quantity of interest involves layers of stochastic dependence, such as in risk measures, stochastic optimal control, high-dimensional filtering, quantum circuit design, or scientific computations involving expensive models. While conceptually natural, NMC presents distinctive algorithmic, computational, and statistical challenges: standard estimators are typically expensive, have suboptimal convergence properties, and in many cases are fundamentally biased for nonlinear targets. A wide range of advanced variants and analysis techniques have been developed, including multilevel strategies, bandit-based tree search, and unbiased or variance-reduced recursive algorithms.

1. Mathematical Formulation and Canonical Algorithms

The archetypal NMC problem is to estimate $\mathbb{E}_Y[f(Y, \gamma(Y))]$, where $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$ and $f$ is typically nonlinear in its second argument. The standard two-level NMC estimator draws $N$ i.i.d. outer samples $Y_n$ and, for each $n$, draws $M$ i.i.d. inner samples $Z_{n,m}$ conditional on $Y_n$, estimating $\gamma(Y_n)$ by a sample mean:

$\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$0

The estimator is unbiased if $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$1 is affine in its second argument, but generally exhibits bias otherwise, even as $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$2 in a fixed ratio. Mean squared error is governed by $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$3, so with $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$4, total root mean square error decays as $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$5 for computational budget $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$6 (Rainforth et al., 2016).

This inability to eliminate bias for nonlinear $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$7 is generic for nested Monte Carlo: Rainforth et al. demonstrate that no estimator based solely on unbiased inner conditional simulations can produce unbiased outer samples for all $\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$8 (Rainforth et al., 2016). This inherent bias, slow convergence, and compounded variance motivate most recent innovation in the field.

2. Nested Monte Carlo in Combinatorial and Optimization Contexts

Nested MCS approaches have been adapted extensively for settings where the feasible solution space or policy configuration admits a recursive or sequential structure, including automated design and optimization.

A salient example is nested Monte Carlo tree search (N-MCTS) and its hybrids, such as the N-MCTS+CMAB framework for quantum circuit synthesis (Wang et al., 2022). In this framework, each node in a search tree encodes a partially constructed solution (e.g., an incomplete quantum circuit), and the search tree is traversed recursively, with each node representing a local combinatorial multi-armed bandit (CMAB) subproblem. The algorithm maintains local reward and visitation statistics, using an upper confidence bound (UCB) principle for action selection:

$\gamma(Y) = \mathbb{E}_Z[\phi(Y,Z) | Y]$9

where $f$0 and $f$1 track visits and rewards. Crucially, rollouts (simulations from a node to a leaf) are themselves performed via (possibly lower-level) N-MCTS, enabling guided exploration in exponentially large spaces. The approach is highly scalable, producing optimized structures in search spaces of cardinality $f$2–$f$3 within $f$4 iterations, outperforming flat or random protocols (Wang et al., 2022).

The same recursive principle underpins search strategies in classical combinatorial optimization, such as NMCS for graph coloring (Cazenave et al., 4 Apr 2025). Here, the search recursively explores move sequences, where level-$f$5 recursions call level-$f$6 searches for each possible next move, selecting continuations that maximize playout returns. While effective in small-branching regimes, CPU cost scales exponentially in nesting depth.

3. Advanced and Multilevel Variants

The poor asymptotic efficiency and deep-nesting bias of vanilla NMC estimators have motivated several algorithmic innovations:

• Unbiased randomized multilayer estimators. The recursive estimator for arbitrary depth (READ) leverages randomization and telescopic sum decompositions for repeated nestings of arbitrary depth $f$7 (Syed et al., 2023). By invoking randomized level differences (via geometric stopping and antithetic sampling structurally akin to Rhee-Glynn or MUSE estimators), the READ approach achieves overall $f$8 computational cost under mild regularity. Unlike classical NMC, the READ estimator is exactly unbiased, trivially parallelizable, and scales without exponential cost in $f$9.
• Multilevel and adaptive MLMC. For problems characterized by a hierarchy of discretization or approximation levels (e.g., SDE and risk estimation), multilevel Monte Carlo (MLMC) and its nested forms (including antithetic and biased-correction variants) recover canonical $N$0 complexity for discontinuous targets under suitable conditions, using telescoping sums across levels (Haji-Ali et al., 2023, Haji-Ali et al., 2021, Giles et al., 2021). Adaptive refinement—where inner simulation effort is increased only when required for accuracy—yields additional speedups (Haji-Ali et al., 2021).
• Variance reduction and regression-based nesting. For pricing of American/Bermudan options and related path-dependent payoffs, the complexity of inner conditional expectation estimation is dramatically reduced by nonparametric regression control variates, yielding complexities below $N$1 under sufficient smoothness (Belomestny et al., 2016).

4. NMC in Computational Physics, Bayesian Inference, and Filtering

In computational physics, NMC schemes are used for efficient equilibrium sampling with expensive "full" potentials, using inexpensive reference models as inner Markov chains to decorrelate configurations and optimize acceptance rates via variance-minimization criteria (Leiding et al., 2013). Optimization of the reference temperature parameter $N$2 on the fly compensates for inaccuracies in the reference, yielding speed-ups by factors up to $N$3 over unoptimized NMC and competitive performance against ab-initio MD.

In Bayesian filtering, nested sequential Monte Carlo (NSMC) methods are implemented for high-dimensional state-space models (Naesseth et al., 2016, Naesseth et al., 2015). NSMC approximates the (generally intractable) fully-adapted proposal distribution by a nested particle filter or importance sampling, yielding "properly weighted" pairs (sample, weight) at each stage. The approach scales SMC to systems with state dimensions $N$4–$N$5, restoring near-ideal resampling properties even in high-dimensional non-Gaussian models.

For Bayesian experimental design and nested integration, double-loop randomized quasi-Monte Carlo (rQMC) estimators are deployed, combining low-discrepancy sampling with specialized scrambling and importance techniques to control both bias and variance—achieving one to two orders of magnitude variance reduction over standard NMC in applied problems (Bartuska et al., 2023).

5. Complexity, Bias, and Trade-offs

Nested Monte Carlo suffers from slow convergence: unless $N$6 is affine, root mean square error (RMSE) for two-level NMC decays as $N$7, requiring cost $N$8 for RMSE less than $N$9, versus $Y_n$0 for ordinary (non-nested) MC (Rainforth et al., 2016, Chen et al., 25 Feb 2025). Deeper nestings compound the cost, with classical NMC and MLMC growing as $Y_n$1 and sometimes $Y_n$2 for $Y_n$3 layers (Syed et al., 2023).

For certain target classes, variance reduction, regression/ML methods, and adaptive multilevel sampling can restore near-optimal or minimax rates. Unbiasedness is generally attainable only for linear or specially structured nonlinearities; general estimation of $Y_n$4 with nonlinear $Y_n$5 will necessarily involve bias unless advanced rMLMC or similar techniques are used (Rainforth et al., 2016, Syed et al., 2023). Trade-offs between sample variance (innate to MC) and computational cost (dominated by the inner simulation loops) are central to design choices.

6. Practical Applications and Performance in Scientific Computing

NMC methods are foundational in quantifying risk measures in finance, American option pricing, Bayesian experimental design, PDE/BSE discretization, recursive planning, and automated quantum circuit design. Key empirical findings include:

• In quantum circuit search, N-MCTS+CMAB enables tractable search over spaces up to $Y_n$6 elements, rapidly identifying near-optimal designs in a handful of iterations and outperforming exhaustive and random alternatives (Wang et al., 2022).
• In high-dimensional PDEs, nesting MC with judicious branching and truncation achieves dimension-independent error decay, with the nested estimator free from the curse of dimensionality (Warin, 2018).
• For high-dimensional Bayesian filters, NSMC remains competitive or superior to alternatives on the basis of effective sample size, statistical efficiency, and cost, scaling to problems with hundreds to thousands of latent variables (Naesseth et al., 2016, Naesseth et al., 2015).
• In ab initio molecular dynamics, NMC with reference-system optimization rescues poor initial models, obtains $Y_n$7 acceptance rates and up to $Y_n$8 speedup, while retaining full Boltzmann sampling (Leiding et al., 2013).

7. Limitations, Open Problems, and Future Directions

NMC is fundamentally limited by its polynomial or exponential overhead in nesting depth and the inescapable bias for general nonlinear targets. Substantial progress has been made in polynomial-time algorithms for iterated nested expectations, notably via multilevel Picard (MLP) schemes exploiting full-history recursion and telescoping differences to reduce the complexity in number of nestings $Y_n$9 from exponential to polynomial in $n$0 and $n$1 (Beck et al., 2020).

Future work is directed toward: (i) Closing remaining gaps in unbiased/debiased estimation of broad classes of nonlinear nested functionals; (ii) Integration of kernel and quasi-Monte-Carlo surrogates; (iii) Efficient parallelization and adaptation to large-scale streaming contexts; and (iv) Theoretical elucidation of optimal nested variance reduction and estimator composition, especially in non-asymptotic regimes and for non-Gaussian, heavy-tailed, or high-dimensional input models.

The Nested Monte Carlo framework thus occupies a central, highly active niche in contemporary stochastic simulation, with applications spanning algorithmic learning, quantum computing, complex physical modeling, and high-dimensional data assimilation.