PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Monte Carlo Methods for Parameter Estimation

D. Luengo, L. Martino, M. Bugallo, V. Elvira, S. Särkkä

arXiv:2107.11820v1stat.COcs.AIeess.SPmath.NA

TL;DR

Analytical estimators are often unavailable for static-parameter inference in real-world signal-processing applications. The paper reviews Monte Carlo methods and tests representative algorithms across benchmark and signal-processing problems, providing a unified resource for practitioners.

  • Problem

    Static-parameter estimation in signal processing often lacks analytically computable ML, MAP, or MMSE estimators.

  • Method

    The paper reviews Monte Carlo methods, including rejection sampling, MCMC algorithms, importance sampling, and their combinations, in a unified format.

  • Results

    The reviewed algorithms are tested on two simple problems, chaotic-system parameter estimation, wireless-sensor-network localization, and spectral analysis of multiple sinusoids.

  • Takeaways & Limitations

    The unified presentation is intended to let signal-processing practitioners apply relevant Monte Carlo algorithms to their specific problems.

  • Takeaways & Limitations

    Rejection sampling can be applied universally only when an upper bound C can be found, and it has practical drawbacks.

Abstract

from arXiv · show

Statistical signal processing applications usually require the estimation of some parameters of interest given a set of observed data. These estimates are typically obtained either by solving a multi-variate optimization problem, as in the maximum likelihood (ML) or maximum a posteriori (MAP) estimators, or by performing a multi-dimensional integration, as in the minimum mean squared error (MMSE) estimators. Unfortunately, analytical expressions for these estimators cannot be found in most real-world applications, and the Monte Carlo (MC) methodology is one feasible approach. MC methods proceed by drawing random samples, either from the desired distribution or from a simpler one, and using them to compute consistent estimators. The most important families of MC algorithms are Markov chain MC (MCMC) and importance sampling (IS). On the one hand, MCMC methods draw samples from a proposal density, building then an ergodic Markov chain whose stationary distribution is the desired distribution by accepting or rejecting those candidate samples as the new state of the chain. On the other hand, IS techniques draw samples from a simple proposal density, and then assign them suitable weights that measure their quality in some appropriate way. In this paper, we perform a thorough review of MC methods for the estimation of static parameters in signal processing applications. A historical note on the development of MC schemes is also provided, followed by the basic MC method and a brief description of the rejection sampling (RS) algorithm, as well as three sections describing many of the most relevant MCMC and IS algorithms, and their combined use.

1 Introduction

Statistical signal processing estimates unknown parameters from noisy observations, but real-world ML, MAP, and MMSE estimators often lack closed forms. This introduction motivates Monte Carlo methods and situates the survey historically and methodologically.

  • Motivation: Signal processing inference estimates unknown parameters from noisy observations in applications including denoising, blind deconvolution, and target localization.Bayesian formulations combine prior information with an input-output likelihood to obtain a posterior distribution.
  • Motivation: Closed-form expressions for ML, MAP, and MMSE estimators are usually unavailable in practical problems, motivating approximate inference methods such as Monte Carlo.The paper notes that computational advances and Monte Carlo development reduce reliance on restricted model classes.
  • Monte Carlo methods: Monte Carlo methods approximate optimization and inference problems by generating random parameter realizations and replacing integrations with sample averages.The approach can use physically replicated experiments or probabilistic simulation to obtain candidate values.
  • Historical development: Monte Carlo methods developed from early stochastic experiments, became a general computational method with 1940s computers, and later produced MH and Gibbs sampling milestones.The historical account spans Buffon, Ulam, Metropolis, Geman and Geman, and subsequent developments.
  • Scope and organization: The survey covers recent advances and presents relevant Monte Carlo algorithms in a unified notation for static-parameter estimation in signal processing.It includes MCMC, importance sampling, adaptive and gradient-based methods, advanced schemes, and their combinations.

2 Mathematical Formulation

The paper formulates Bayesian parameter estimation as posterior integration and introduces Monte Carlo and rejection sampling approximations. Direct posterior sampling is often unavailable, while rejection sampling additionally requires a usable envelope and can be inefficient.

  • 2.1 Bayesian Inference and the Monte Carlo Method: A dataset y depends on a static parameter vector θ, and Bayesian inference represents information about θ through the posterior or target PDF.The posterior is expressed using a likelihood, prior, and marginal-likelihood normalization term.
  • 2.1 Bayesian Inference and the Monte Carlo Method: The desired integral of g(θ) under the posterior includes MMSE estimation when g(θ) = θ.The MMSE estimator is the conditional expectation of θ with respect to the posterior PDF.
  • 2.1 Bayesian Inference and the Monte Carlo Method: The basic Monte Carlo estimator uses IID posterior samples and an unweighted sample average to approximate the integral.The estimator is unbiased and converges almost surely by the strong law of large numbers; a CLT applies when g is square integrable.
  • 2.1 Bayesian Inference and the Monte Carlo Method: When direct posterior sampling is impossible but point-wise target evaluations are available, rejection sampling, MCMC, and importance sampling provide alternatives.These algorithm families form the core of the paper.
  • 2.2 Rejection Sampling (RS): Rejection sampling draws from a simpler proposal and accepts candidates using a PDF-ratio test under an envelope condition Cq(θ) ≥ π(θ).Accepted samples follow the target density, and the procedure stops after obtaining the required number of valid samples.
  • 2.2 Rejection Sampling (RS): Rejection sampling can be difficult in high dimensions because finding the envelope bound may be hard and acceptance probability may be low.Low acceptance probability can render the method practically useless for some target-proposal pairs.
  • 2.2 Rejection Sampling (RS): The number of iterations needed to generate M accepted samples is random, with expected value E(T) = M/PA, preventing exact generation time from being fixed in advance.This unpredictability can be serious in practical applications.
  • 2.2 Rejection Sampling (RS): Adaptive rejection sampling addresses some difficulties, but envelope construction remains restrictive in high-dimensional spaces, where MCMC and importance sampling are generally more efficient.The paper identifies this as a boundary on generalized adaptive RS applicability.

3 Markov chain Monte Carlo (MCMC)

MCMC methods generate an ergodic Markov chain whose stationary density is the desired target PDF. The survey proceeds from basic MH and Gibbs samplers to adaptive, gradient-based, and other advanced schemes.

  • MCMC definition: An MCMC method produces an ergodic Markov chain with the desired target PDF as its stationary density.The survey uses this definition to organize its discussion of MCMC algorithms.
  • MCMC scope: The MCMC discussion begins with MH and Gibbs samplers and continues to adaptive, gradient-based, and other advanced schemes.These sections cover the principal algorithmic building blocks and extensions reviewed in the paper.

3.1 MCMC basic building blocks: Metropolis-Hastings and Gibbs samplers

MH and Gibbs samplers provide the survey’s basic MCMC building blocks. MH combines proposal moves with acceptance decisions, while Gibbs updates coordinates from full conditionals; proposal tuning and coordinate updating govern convergence and efficiency.

  • Metropolis-Hastings: The MH algorithm accepts or rejects proposal-based particle moves according to energy changes, retaining rejected states for subsequent averages.Its origins lie in sampling molecular configurations according to the Boltzmann distribution.
  • Metropolis-Hastings: MH proposals depend on the previous state, so samples are correlated, but the chain can target distributions that lie below the proposal density.Reusing the previous state after rejection creates dependence while preserving the desired limiting invariant distribution.
  • Metropolis-Hastings: Hastings generalized Metropolis and Barker acceptance rules to non-symmetric proposals, and Peskun later established the Metropolis rule’s optimality.The resulting acceptance structure is the one commonly used today.
  • Metropolis-Hastings: The MH burn-in period removes initial samples from empirical averages to allow approximate convergence toward the stationary distribution.The algorithm specifies a proposal, initial state, total iterations, and burn-in period before forming the estimate.
  • Metropolis-Hastings: MH is generic across many proposal and target PDFs, but convergence speed and estimator quality depend substantially on proposal quality.This makes proposal design a central practical issue despite the algorithm’s broad validity.
  • Metropolis-Hastings: MH acceptance rates should avoid both extremes: high proposal variance causes rejection and low variance can prevent exploration of the target.For random-walk MH, reported optimal average acceptance rates range from approximately 44% in one dimension to 23% as dimension grows.
  • Metropolis-Hastings: A small-world mixture proposal combines local and global moves, improving performance in heterogeneous or multimodal spaces and potentially turning slow mixing into rapid mixing.The global component enables distant moves with small probability, while the local component remains centered near the current state.
  • Gibbs sampler: Gibbs sampling updates one coordinate at a time from its full conditional, accepting each proposed coordinate value automatically.Ergodicity requires that every coordinate continue to be visited infinitely often.

3.1.4 MH-within-Gibbs

MH-within-Gibbs embeds Metropolis–Hastings updates inside Gibbs sampling to handle non-standard full conditional PDFs without requiring rejection-sampling bounds. Griddy Gibbs and FUSS provide proposal constructions, with FUSS typically yielding high acceptance and low correlation.

  • Motivation: Gibbs sampling can require full conditional PDFs that have non-standard forms and cannot be sampled directly.Rejection sampling was initially used to address this difficulty.
  • Griddy Gibbs: Griddy Gibbs evaluates the conditional target on a grid, builds a piecewise inverse-CDF approximation, and draws samples by inversion.The approximation may use piecewise constant or piecewise linear interpolation.
  • Griddy Gibbs: Griddy Gibbs samples are only approximately distributed according to the target, while constructing multivariate inverse-CDF approximations remains challenging.The Gibbs stopper can correct the first issue with importance weights, but not the multivariate approximation challenge.
  • MH-within-Gibbs: MH-within-Gibbs uses an internal MH algorithm to sample conditionals that cannot be sampled directly and can also operate within block or collapsed Gibbs samplers.The internal MH step need not be used for parameters whose conditionals are directly sampleable.
  • FUSS within Gibbs: FUSS within Gibbs prunes a dense support grid, constructs a piecewise proposal, and applies internal MH steps using that proposal.Its target-tailored proposal usually produces very high acceptance, low sample correlation, and reduced estimator variance, especially for peaky proposals.

3.1.5 Other Classical MCMC Techniques

Other classical MCMC techniques address difficult multivariate targets through augmentation, geometric transformations, adaptive directions, and chain thinning. These methods introduce useful alternatives but retain computational or variance trade-offs.

  • Data augmentation: Data augmentation introduces latent data so Bayesian posterior analysis becomes easier, paralleling the basic idea of expectation-maximization.The method can be implemented as block or component-wise Gibbs sampling.
  • Slice sampling: Slice sampling represents the target through the region below its unnormalized density and samples uniformly from that augmented region.The auxiliary variable makes the target coordinate marginally distributed according to the desired density.
  • Slice sampling: Slice sampling can be impractical because determining the slice set requires inverting a usually non-monotonic target, often producing disjoint sets.This difficulty can arise even for unidimensional targets.
  • Hit-and-run and adaptive directions: Hit-and-run methods explore parameter spaces with random directions, while adaptive direction sampling maintains and updates support points to determine sampling directions.Hit-and-run and random-scan Gibbs arise as special cases of the general adaptive direction framework.
  • Hit-and-run and adaptive directions: Direct sampling along arbitrary directions may be infeasible, motivating Metropolised hit-and-run variants.The Metropolised version addresses cases where direct sampling from the target along a direction is unavailable.
  • Thinning: Thinning discards K −1 of every K Markov-chain outputs to reduce correlation and computational or storage burden, but commonly increases estimator variance.It can nevertheless improve statistical efficiency when sampled-quantity computation is costly and autocorrelations decay slowly.

3.2 Adaptive MCMC

Adaptive MCMC improves proposal tuning by learning target geometry or adapting one-dimensional proposals. The survey covers covariance and acceptance-rate adaptation, ARMS, and IA2RMS.

  • Adaptive MCMC: Adaptive MCMC addresses the crucial MH proposal-selection problem by using previous samples to improve proposal distributions.Adaptive methods can be especially useful when direct high-dimensional parametric tuning is difficult.
  • Adaptive Metropolis: Adaptive Metropolis estimates the target covariance empirically and adapts the proposal scale toward a target acceptance rate.A small positive constant prevents the covariance estimate from becoming ill-conditioned.
  • Adaptive Metropolis: Under idealized Gaussian conditions, the optimal proposal scale is λ*=2.382/Dθ, while the commonly targeted acceptance rate is α*=0.234.The covariance identity is specific to Gaussian target PDFs, so adapting the scale is often desirable.
  • Adaptive Metropolis: Adding empirical covariance and scale updates can substantially improve the proposal relative to its initial form and thereby improve MH performance.The general adaptive Metropolis algorithm introduces these two updates after the acceptance step.
  • ARMS: ARMS combines adaptive rejection sampling with MH to sample virtually from any univariate PDF and can be embedded within Gibbs sampling.Its proposal is built from support points and updated when rejected proposals reveal regions where the proposal is below the target.
  • IA2RMS: IA2RMS adds rejected MH-stage samples to the support set with nonzero probability, allowing proposal adaptation across all regions and ensuring qm(θ) →π(θ) as t →∞.This also decouples proposal construction from algorithm evolution and permits simpler proposals than ARMS.
  • Adaptive MCMC: Adaptive schemes should use finite adaptation or adapt increasingly less often because diminishing adaptation supports convergence results such as a Central Limit Theorem.The survey notes this recommendation in the context of adaptive MCMC chains.

3.3 Gradient-based techniques

Gradient-based MCMC uses log-posterior gradients to propose longer moves while maintaining acceptance. MALA and HMC improve scaling, but numerical integration and parameterization remain practical concerns.

  • Gradient-based MCMC: Gradient-based MCMC uses ∇log π(θ) to form proposals that can make longer jumps without excessively reducing acceptance.This targets improved efficiency in difficult, high-dimensional distributions.
  • MALA: MALA uses a Langevin SDE as an MH proposal, typically approximating the diffusion with one Euler–Maruyama step.The resulting proposal is corrected by the MH acceptance rule.
  • MALA: MALA is limited to local moves because multiple SDE integration steps make the transition-density evaluation intractable.Consequently, its improvement over random-walk MH can be small.
  • MALA: MALA’s asymptotically optimal acceptance rate is approximately 0.574, higher than the roughly 0.234 rate associated with random-walk MH.The higher rate reflects the additional target information supplied through the SDE.
  • HMC: HMC uses Hamiltonian trajectories involving position and momentum variables to construct MH proposals with extended movement through the state space.Leapfrog integration preserves volume and reversibility needed for a valid MH method.
  • HMC: HMC requires O(d1/4) steps to traverse the state space asymptotically and has an optimal acceptance rate of 0.651.The survey reports this rate as higher than MALA’s 0.574.
  • Riemann-manifold methods: Riemann-manifold methods address parameterization sensitivity by performing Langevin or Hamiltonian simulations on a metric geometry defined by G(θ).The metric generalizes Euclidean distance through d2=dθ^TG^-1(θ)dθ.

3.4 Other advanced schemes

Advanced MCMC schemes address slow or trapped exploration, parallel-computation needs, and extensions beyond reversible chains while preserving the target distribution under appropriate conditions.

  • Parallel MCMC chains can improve state-space exploration when a single chain is trapped in a local mode or converges slowly.This is especially relevant for multimodal posteriors and high-dimensional applications.
  • Parallelization can reduce MCMC computation time and support big-data methods that split a posterior into partial sub-posteriors.The best strategy depends on the hardware: GPU methods may parallelize individual steps, whereas distributed systems may parallelize the algorithm itself.
  • Delayed Rejection Metropolis: Delayed Rejection Metropolis draws a second candidate from a proposal conditioned on the rejected first candidate.Its acceptance function is designed to preserve ergodicity and detailed balance; further candidates can be proposed sequentially after additional rejections.
  • Non-reversible chains: Non-reversible MH-type chains can retain the desired invariant density while avoiding reversibility.Their practical improvement in complex problems remains uncertain.

3.5 MCMC Convergence Diagnostics

MCMC convergence diagnostics assess whether finite chains represent the target distribution and whether their correlation reduces sampling efficiency. Recommended practice combines multiple-chain checks, PSRF monitoring, debugging, and awareness of newer but immature diagnostics.

  • Finite MCMC runs may fail to converge, producing samples that poorly represent the target density and erroneous inference.The transitory period required to reach stationarity is theoretically difficult to determine.
  • Recommended diagnostics run at least three chains from varying starting points, retain their second halves, and assess between-chain and within-chain mixing.After approximate convergence, the retained halves are combined to summarize the target distribution.
  • PSRF: The PSRF should approach 1 at convergence; values significantly above 1 indicate that convergence probably has not occurred.PSRF can be computed from multiple chains or from splits of a single chain.
  • Effective Sample Size: Positive sample correlation lowers effective sample size below the total number of iterations and increases estimator variance relative to independent sampling.With zero correlation, ESS equals the number of iterations, as in classical Monte Carlo.
  • Other Recent Approaches: Stein discrepancies and computational-complexity analyses offer newer convergence or efficiency assessments, but these approaches are not yet mature for widespread practical use.Some Stein discrepancy measures use linear programs or closed-form kernel evaluations.

4 Importance Sampling

Importance sampling approximates target distributions and their moments by retaining all proposal samples and weighting them according to their quality. The section covers single and multiple proposals, adaptive updates, estimator convergence, variance, and combinations with resampling or MCMC.

  • 4 Importance Sampling: IS retains every sample drawn from one or more proposal PDFs and assigns weights based on how well samples represent the target distribution.Standard IS uses one proposal, whereas multiple IS uses a collection of proposals.
  • 4.1.1 Importance sampling with a single proposal: A single proposal should have heavier tails than the target because faster-decaying proposal tails can produce infinite estimator variance.This condition avoids a failure mode caused by the proposal density appearing in the denominator of importance weights.
  • 4.1.2 Importance sampling with multiple proposals: Multiple proposals improve multimodal fitting and parameter-space exploration when selecting one suitable proposal a priori is difficult.This framework is called multiple importance sampling and underlies many adaptive IS algorithms.
  • 4.1.2 Importance sampling with multiple proposals: Deterministic-mixture weighting can reduce variance relative to standard MIS, but it requires N^2 proposal evaluations instead of N.Both weighting alternatives require the same number of target evaluations.
  • 4.2 Adaptive importance sampling: Adaptive IS iteratively updates proposal parameters, while scale adaptation can create ill-conditioned proposals with potentially infinite estimator variance.The review focuses on location-scale proposals and notes that many algorithms adapt locations while keeping scales fixed.
  • 4.3 Convergence of IS and AIS methods: For fixed proposals satisfying q(θ) > 0 whenever π(θ) > 0, unnormalized IS estimates are unbiased and self-normalized estimates converge asymptotically.These guarantees hold for both standard-MIS and deterministic-mixture weights; convergence rates can still differ substantially.

5 MC-within-MCMC methods

Hybrid MC-within-MCMC methods remain MCMC algorithms because they use a Markov chain to sample from the target density while relying on inner MC estimators for candidate generation or related tasks.

  • Hybrid MC-within-MCMC methods use inner Monte Carlo estimators at each MCMC iteration.They remain MCMC algorithms because a Markov chain samples from the target probability density function.

5.1 MCMC with multiple candidates for the estimation of a static parameter

Multiple-candidate MCMC methods generate and evaluate several proposals per iteration, with variants using auxiliary samples, independent proposals, or recycled weighted samples. These designs preserve ergodicity through appropriate acceptance mechanisms while trading additional computation for candidate selection and estimation.

  • Generalized MH schemes test several candidates per iteration to increase candidate acceptance rates while preserving chain ergodicity.The extended acceptance probability must be designed appropriately to guarantee ergodicity.
  • Multiple Try Metropolis draws N candidates, selects one using suitable weights, and then accepts or rejects it with a generalized probability.The standard scheme also generates N −1 auxiliary samples when the proposal depends on the previous chain state.
  • Standard Multiple Try Metropolis requires 2N −1 samples per iteration, although only N −1 auxiliary samples enter the acceptance probability.The generation of auxiliary points therefore creates a computational cost in the standard method.
  • Independent Multiple Try Metropolis avoids auxiliary-sample generation when the proposal is independent of the previous state by recycling already drawn candidates.This reduces the need for the auxiliary-sampling step without jeopardizing ergodicity.
  • Group Metropolis Sampling recycles auxiliary weighted samples to form a final Monte Carlo estimator rather than applying resampling at every iteration.Its output is a sequence of weighted-sample sets whose samples jointly approximate the target distribution and its moments.

5.2 MCMC with multiple candidates for the estimation of a dynamic parameter

For dynamic parameters, particle-based MCMC methods exploit target-density factorization to generate candidate paths and combine them with MH or Gibbs updates. PMH and PMMH use particle approximations and marginal-likelihood estimates, while Particle Gibbs preserves dependence on the previous path.

  • PMH exploits factorization of the target density to construct a sequential particle-based proposal for dynamic parameters.Particle weights and a marginal-likelihood estimate support the particle approximation used in the MH step.
  • Particle Metropolis-Hastings runs a particle filter, selects a weighted particle, and accepts or rejects it using an MH probability involving marginal-likelihood estimators.The method is designed for dynamic variables such as hidden states in state-space models.
  • When no resampling is applied, I-MTM2 and PMH coincide, differing only in whether candidates are generated in batch or sequentially.This establishes a direct relationship between multiple-try and particle-based MH constructions.
  • Particle Marginal Metropolis-Hastings extends PMH to jointly infer dynamic states x and static parameters λ.For a proposed λ, a particle filter approximates the state distribution and the marginal likelihood p(y|λ).
  • Particle Gibbs uses a conditional particle filter incorporating the previous path and is guaranteed to produce a chain with the joint posterior as invariant density.The conditional particle filter is run given the previous static parameter and reference path before selecting the next path.

5.3 Pseudo-marginal MCMC methods

Pseudo-marginal MCMC addresses targets or likelihoods that cannot be evaluated exactly by replacing them with suitable Monte Carlo estimates or simulated data. The reviewed methods include PMMH, exchange methods, and ABC, with estimator variance and simulation requirements defining important practical boundaries.

  • Pseudo-marginal MCMC targets a marginal distribution that cannot be evaluated directly by using an unbiased Monte Carlo estimator of that target.The estimator can be supplied by another Monte Carlo method such as importance sampling.
  • Pseudo-marginal methods require likelihood estimators that are unbiased and valid for every possible parameter value.Importance sampling is commonly used to construct the unbiased likelihood estimate.
  • Pseudo-marginal methods can produce very slowly mixing chains when the unbiased estimator has high variance, despite asymptotic convergence.The review therefore emphasizes the importance of geometric convergence for avoiding arbitrarily slow rates.
  • Single Variable Exchange approximates an intractable likelihood normalizing constant by drawing auxiliary data from the likelihood and embedding the estimate in an MH target.The method is applicable when samples can be drawn from the likelihood even though part of its evaluation is unavailable.
  • ABC avoids likelihood evaluation by simulating artificial data and weighting parameter simulations according to their distance from observed data.As ϵ approaches zero, the generalized ABC target becomes increasingly similar to the posterior target.
  • For large datasets, adaptive subsampling uses cheaper likelihood evaluations based on random data subsets while controlling deviation from the desired target.This addresses settings where full-data likelihood evaluations are prohibitively expensive.

5.4 Noisy and approximate likelihood methods

When the likelihood cannot be evaluated exactly, noisy or approximate likelihoods can be incorporated into MH by estimating the acceptance probability or likelihood. Unbiased likelihood estimates preserve validity, whereas approximate MCMC requires careful analysis because small chain perturbations may alter stationarity substantially.

  • Noisy likelihood methods: The unified framework replaces the standard MH acceptance probability with an approximation when it cannot be evaluated directly.The approximation is constructed using a proposed parameter and an auxiliary draw from the likelihood model.
  • Noisy likelihood methods: A randomized acceptance-probability estimate can yield a chain whose stationary density approximates the desired posterior PDF.The estimate is formed using y′ ∼ ℓ(y|θ′) and should satisfy the stated condition; its deviation should be small for a useful approximation.
  • Approximate likelihood methods: In state-space models, approximate likelihoods are useful when nonlinear dynamics prevent exact marginal-likelihood evaluation.Nonlinear Kalman filter-based Gaussian approximations can provide an available approximation to the likelihood.
  • Approximate likelihood methods: Nonlinear Kalman filter likelihood estimates are generally biased, but studies report good approximations to the true posterior in discrete-time and stochastic-differential-equation models.The cited filters include extended, unscented, and cubature Kalman filters.
  • Analysis of noisy/approximate likelihood methods: A Markov chain close to another need not have a stationary distribution, and if it does, that distribution need not be close to the original chain’s stationary distribution.This makes the analysis of noisy or approximate MCMC delicate and motivates perturbation bounds for uniformly and geometrically ergodic chains.

6 Numerical simulations

The simulations evaluate Monte Carlo algorithms across Gaussian, adaptive importance-sampling, chaotic-system, and wireless-sensor-network examples, showing that tailored proposals and adaptive designs can substantially improve estimation and exploration.

  • 6 Numerical simulations: The experiments begin with analytically tractable univariate and bivariate Gaussian targets, then examine chaotic-system estimation and wireless-sensor-network localization.These examples allow exact performance assessment in the Gaussian cases and evaluation on more challenging signal-processing problems.
  • 6.1 Illustrative Example for Adaptive MCMC algorithms: The multimodal Gaussian example shows AGM-MH achieving an MSE two orders of magnitude lower than AM and autocorrelation up to 2.5 times smaller.AGM-MH uses a Gaussian-mixture proposal, while AM uses an adaptive random-walk proposal.
  • 6.1 Illustrative Example for Adaptive MCMC algorithms: AGM-MH frequently explores all three modes, whereas random-walk MH with σ = 2 completely misses one mode and σ = 5 covers all three.The result illustrates the sensitivity of random-walk MH to proposal variance.
  • 6.2 Illustrative example for Adaptive Importance Sampling: For adaptive importance sampling, compared schemes outperform standard PMC across σ, with local resampling generally better than global resampling.APIS performs well for intermediate σ, AMIS for large σ, and the best K depends on the proposal scale.
  • 6.3 Parameter estimation in a chaotic system: FUSS-within-Gibbs attains very small MSEs across λ and outperforms MH by tailoring proposals to narrow full-conditionals.This remains effective even when the conditional density has a sharp peak resembling a delta function.
  • 6.5 Spectral analysis: With fixed posterior-evaluation effort E = MT = 10^4, O-MCMC achieves lower relative errors than independent parallel chains, with similar performance as effort grows.The comparison varies the number of chains and epochs while keeping computational effort fixed.

7 Conclusion

The paper reviews Monte Carlo methods for static-parameter estimation in signal processing, emphasizing a unified presentation of methods that address analytically intractable estimators. It also tests these algorithms across illustrative, scientific, and signal-processing problems while cautioning that improper application can yield infinite-variance estimators.

  • 7 Conclusion: The review presents Monte Carlo methods for approximate inference of static parameters when analytical estimators cannot be computed.It focuses on relevant methods rather than specific applications.
  • 7 Conclusion: Many algorithms are presented in a clear, unified format intended for direct application by signal processing practitioners.
  • 7 Conclusion: The paper develops a self-contained treatment from basic Monte Carlo methods and convergence properties through classical and advanced MCMC algorithms.It also includes importance sampling and adaptive importance sampling for completeness.
  • 7 Conclusion: The algorithms are evaluated on simple problems, chaotic-system parameter estimation, wireless-sensor-network localization, and spectral analysis of multiple sinusoids.
  • 7 Conclusion: Monte Carlo methods can produce infinite-variance estimators when applied improperly, as illustrated by the weighted likelihood bootstrap example.Although the cited approach is asymptotically unbiased, its infinite variance can produce estimates far from the correct values.
Loading 2107.11820v1…