SMC–WFR: Weighted Forward Recursion

• SMC–WFR is a family of sequential Monte Carlo methods that use weighted forward recursion to control variance, mitigate path degeneracy, and support online smoothing.
• It integrates advanced resampling techniques and weight-function resets to efficiently approximate filtering distributions and additive smoothing expectations.
• Recent extensions connect SMC–WFR to Wasserstein–Fisher–Rao gradient flows, achieving exponential KL contraction and improved performance in rare-event simulation.

Sequential Monte Carlo (SMC–WFR) encompasses a family of methodologies within sequential Monte Carlo that utilize weighted forward recursions, advanced resampling algorithms, or weight-function resets to achieve variance control, robustness to path degeneracy, and efficiency in high-dimensional or rare-event inference. The abbreviation “SMC–WFR” is variously used in the literature for “Weighted Forward Recursion” (Moral et al., 2010), “Weight-Function Resetting” (Naesseth et al., 2019), and “Weighted Finite Resampling” (Webber, 2019). Recent developments also connect SMC–WFR to Wasserstein–Fisher–Rao (WFR) gradient flows for optimization in measure spaces (Crucinio et al., 6 Jun 2025). SMC–WFR enables online smoothing, sharp asymptotic control of estimator variance, and, in specialized forms, logarithmically efficient importance sampling for large-deviation problems (Chan et al., 2012). This article provides a comprehensive account of SMC–WFR from formal models and algorithms to advanced theoretical properties and variants.

1. State-Space Setting and Smoothing via Weighted Forward Recursion

Consider a hidden Markov model (HMM) on a state-space $\mathcal{X}$ with observations $Y_k$ in $\mathcal{Y}$. The joint density, under parameter $\theta$, has the form

$$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$

with $X_0\sim \mu_\theta(\cdot)$, $X_k|X_{k-1}=x_{k-1}\sim f_\theta(\cdot|x_{k-1})$, and $Y_k|X_k=x_k\sim g_\theta(\cdot|x_k)$.

The central inferential objective is the recursive computation of smoothed expectations of additive path functionals: $$S_n(x_{0:n}) = \sum_{k=1}^n s_k(x_{k-1},x_k), \qquad \mathcal{S}_n^\theta = \mathbb{E}_\theta\left[ S_n(X_{0:n}) \mid y_{0:n} \right]$$ A forward-smoothing (weighted forward recursion, WFR) enables one to compute $\mathcal{S}_n^\theta$ recursively in $Y_k$0 without path storage, through auxiliary functions

$Y_k$1

satisfying

$Y_k$2

with $Y_k$3 (Moral et al., 2010).

2. The SMC–WFR Algorithm: Structure, Recursion, and Pseudocode

In practice, the filtering distribution $Y_k$4 and backward kernel $Y_k$5 are not available in closed form. SMC–WFR substitutes empirical measures constructed from a weighted particle cloud:

• At time $Y_k$6, maintain particles $Y_k$7 and corresponding estimates $Y_k$8.
• Propagate and reweight to obtain $Y_k$9 approximating $\mathcal{Y}$0.
• For each $\mathcal{Y}$1, update the forward-smoothing estimate by

$\mathcal{Y}$2

• The smoothed additive expectation is then

$\mathcal{Y}$3

This results in a fully online algorithm, summarized as follows:

$$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$6 Under this scheme, the key smoothing weights are

$\mathcal{Y}$4

(Moral et al., 2010).

3. Theoretical Properties: Variance, Consistency, and Central Limit Theorems

SMC–WFR is designed to control the variance of smoothed estimators and mitigate path degeneracy:

• Mean-square error: For $\mathcal{Y}$5 and all $\mathcal{Y}$6,

$\mathcal{Y}$7

where $\mathcal{Y}$8 is independent of $\mathcal{Y}$9.

• Central Limit Theorem: As $\theta$0,

$\theta$1

with $\theta$2 growing at most linearly with $\theta$3. This is in marked contrast with path-space particle smoothers, for which the variance increases at least quadratically in $\theta$4 (Moral et al., 2010).

• The analysis is based on telescoping the estimation error into one-step increments, using contractive properties of the forward and backward kernels and Khinchine inequalities for each step.

SMC–WFR is proven to be unbiased for functions of the terminal state and establishes consistency and CLT properties for a wide function class, with variance bounded by the (at most) linear-in-time accumulation due to resampling noise (Rohrbach et al., 2022).

4. Algorithmic Variants: Weight-Function Resetting and Matrix-Resampling

The WFR principle applies beyond forward-smoothing. In SMC with Weight-Function Resetting (Naesseth et al., 2019), blocks of SMC resample and normalize weights, controlling early path degeneracy by periodically setting incremental weights to unity:

• Fix reset times $\theta$5, either uniformly or adaptively (e.g., ESS-triggered).
• On block $\theta$6 run standard SMC, initializing and resetting weights at the start of each block.

Table: Distinct SMC–WFR Algorithms and Their Key Elements

Variant	Main Mechanism	Noted Effect/Strength
Weighted Forward Rec.	Forward additive update	Linear-in-time variance for smoothing
Weight-Function Reset	Blockwise weight resets	Controls path degeneracy, robust Z-est.
Weighted Finite/Residual	Residual resampling	Variance reduction over multinomial

Matrix-resampling (Webber, 2019) encapsulates WFR as “residual” resampling and identifies optimal schemes:

• Weighted finite (residual): allocate integer copies by floor of normalized weights, then use multinomial allocation for fractional parts.
• Stratified (sorted): sorts by an informative statistic, then stratifies, achieving minimal resampling variance for that statistic.

5. Extensions: Logarithmic Efficiency, Random-Weight SMC, and Rare-Event Estimation

WFR schemes are central to rare-event simulation and SMC with unbiased random weights:

• Logarithmic Efficiency: For estimation of small probabilities (e.g., rare events in Markov additive processes), SMC–WFR resampling weights are constructed to mimic the incremental likelihood ratio of an optimal exponential tilt. This achieves a variance-to-square mean ratio of order $\theta$7 for event probabilities $\theta$8 (Chan et al., 2012).
• Random-Weight SMC: The estimator is consistent in probability under minimal $\theta$9 moment conditions, and a CLT holds for functionals in a recursively defined $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$0 function class; resampling ensures the variance is a sum over steps, rather than a multiplicative product, preventing exponential variance growth (Rohrbach et al., 2022).

Such SMC–WFR estimators play a critical role in high-dimensional and rare-event simulations, where classical importance sampling would otherwise fail due to poor proposal adaptation or intractable optimal tilting.

6. SMC–WFR in Wasserstein–Fisher–Rao Gradient Flows

Recent work (Crucinio et al., 6 Jun 2025) establishes a distinct but nomenclature-overlapping SMC–WFR method to approximate Wasserstein–Fisher–Rao gradient flows in measure space optimization:

• The WFR distance on $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$1 interpolates between 2-Wasserstein (mass-conserving) and Fisher–Rao (mass-varying) transports.
• The WFR gradient flow of $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$2 for sampling is discretized by splitting into a Fokker–Planck (Wasserstein) step and a pure Fisher–Rao (replicator) step.
• Each iteration computes an unadjusted Langevin move (diffusion) followed by importance re-weighting, then resampling, corresponding to the forward-only SMC paradigm.
• The method achieves exponential-rate KL contraction under log-Sobolev assumptions, and numerically outperforms competing schemes on multimodal targets.

7. Computational and Practical Considerations

• Complexity: The basic SMC–WFR (forward-smoothing) update involves $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$3 computation per timestep, reducible to $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$4 with specialized summation/data structures (Moral et al., 2010).
• Memory: Only current and previous particle clouds and associated forward variables are required, yielding $$p_\theta(x_{0:n},y_{0:n}) = \mu_\theta(x_0)\prod_{k=1}^n f_\theta(x_k|x_{k-1}) \prod_{k=0}^n g_\theta(y_k|x_k)$$5 storage overhead.
• Parallelism and Implementation: SMC–WFR variants such as blockwise reset and chain-autocovariance-based variance estimation are naturally parallelizable and suitable for modern architectures (Dau et al., 2020, Naesseth et al., 2019).
• Variance Control: All WFR-family methods are constructed to suppress variance, either by summary-weight propagation (forward recursion), tailored resampling (matrix schemes), blockwise resets, or use of rare-event tilting information.

• "Forward Smoothing using Sequential Monte Carlo" (Moral et al., 2010)
• "Elements of Sequential Monte Carlo" (Naesseth et al., 2019)
• "Unifying Sequential Monte Carlo with Resampling Matrices" (Webber, 2019)
• "Convergence of random-weight sequential Monte Carlo methods" (Rohrbach et al., 2022)
• "A sequential Monte Carlo approach to computing tail probabilities in stochastic models" (Chan et al., 2012)
• "Sequential Monte Carlo approximations of Wasserstein--Fisher--Rao gradient flows" (Crucinio et al., 6 Jun 2025)
• "Waste-free Sequential Monte Carlo" (Dau et al., 2020)