---
title: Adaptive Resampling (ARS) Overview
url: https://www.emergentmind.com/topics/adaptive-resampling-ars
type: topic
---

# Adaptive Resampling (ARS) Overview

Searching arXiv for recent papers and context on adaptive resampling and related ARS terminology.
Adaptive resampling is a family of procedures in which the resampling operation is made data dependent rather than fixed in advance. Across the arXiv literature, this includes particle-filter rules that trigger resampling only when weight degeneracy becomes severe, particle-allocation schemes that expand the population when an intermediate distribution is poorly approximated, image-processing methods that predict spatially varying interpolation kernels and apply them adaptively to each position in the image, and optimization procedures that allocate additional evaluations to uncertain or potentially Pareto-relevant points. The acronym is also overloaded: in neighboring Monte Carlo literature, “ARS” can denote adaptive rejection sampling, which is distinct from adaptive resampling proper [1306.3172] [1503.00966] [1712.06463] [2503.21495] [1509.07985].

## 1. Terminology and conceptual scope

In the literature summarized here, adaptive resampling is not a single canonical algorithm. It is a recurring design principle in which a resampling decision is conditioned on a diagnostic such as ESS, KLD, posterior complexity, class-wise performance, local image structure, or bootstrap-estimated uncertainty. This suggests that “adaptive resampling” is best understood as a methodological pattern rather than a field-specific primitive.

| Domain | What is adapted | Signal used |
|---|---|---|
| Particle filtering | Whether to trigger resampling or how many particles to resample | ESS or KLD [2606.25603] [1306.3172] |
| Resample-Move SMC | Particle-set size at difficult intermediate distributions | $\mathrm{ESS}_n / R_n$ [1604.01972] |
| Image resampling | Spatially varying interpolation kernels | Local content, per-pixel context, or channel-specific structure [1712.06463] [2511.01620] |
| Noisy optimization | Additional evaluations of candidate points | Probability of dominance via bootstrap [2503.21495] |
| Imbalanced classification | Sampling distribution over classes during training | Class-wise macro F1 [2509.00955] |

A recurrent misconception is that adaptive resampling always means “resample more often.” In several formulations the opposite is central: resampling is postponed until needed, or the sampling budget is concentrated only where the approximation is weak. Another misconception is that all ARS methods optimize the same objective. In fact, the objectives range from controlling particle degeneracy and divergence from the target distribution to preserving image detail, reducing variance of $\log \hat Z$, or improving macro-F1.

## 2. Particle filters: KLD-resampling and adaptive sample size

In particle filtering, adaptive resampling can mean determining the number of particles to resample so that the Kullback-Leibler distance between the distribution of particles before resampling and after resampling does not exceed a pre-specified error bound [1306.3172]. The setting is a weighted particle approximation $\{x_t^{(i)}, w_t^{(i)}\}$ to the posterior $p(x_t \mid y_{1:t})$, followed by resampling to an equally weighted set. The method bins the state space into $k$ occupied cells and chooses the resampled size online so that, with probability at least $1-\delta$, the KL divergence remains below $\varepsilon$.

Using the $\chi^2$ approximation and the Wilson-Hilferty approximation, the required sample size is written as
$$
N \ge \frac{k-1}{2\varepsilon}
\left[
1 - \frac{2}{9(k-1)} + \sqrt{\frac{2}{9(k-1)}}\, z_{1-\delta}
\right]^3.
$$
Operationally, the algorithm draws resampled particles one by one, updates the number of occupied bins, and recomputes the required $N$ whenever a new bin becomes non-empty. This differs from Fox’s KLD-sampling because the bound is inserted into the resampling step rather than the sampling step; the formulation therefore measures the fit of the distribution represented by weighted particles based on KLD during resampling rather than in sampling [1306.3172].

The reported target-tracking simulations use a maneuvering-target scenario with RMSE, average number of particles resampled over time, and wall-clock computational time as metrics. The observations are that ARS via KLD-resampling achieves RMSE comparable to both fixed-$N$ SIR-PF and Fox’s KLD-sampling PF, uses significantly fewer particles on average when the posterior is sharp, and automatically increases $N$ when the posterior spreads out during maneuvers. The same summary reports up to 40–50% savings in resampling cost while delivering the same estimation quality at lower average computational cost [1306.3172].

## 3. ESS-triggered adaptive resampling, $\infty$-ESS, and conditional SMC

A second major line of work defines adaptive resampling as an online decision rule based on particle degeneracy. In the simplest ESS-based formulation, one fixes a threshold fraction $\alpha \in (0,1]$ and triggers resampling when
$$
\mathrm{ESS}_t = \frac{1}{\sum_{n=1}^N (\bar w_t^n)^2} < \alpha N.
$$
A smaller $\alpha$ leads to fewer resampling steps, whereas a larger $\alpha$ keeps weights more balanced [2606.25603]. In the unified CSMC framework, adaptive resampling is represented as a mixture of trivial resampling and a proper resampling scheme, with the mixture determined by the ESS criterion. Under the assumptions of proper weighting and persistence, the resulting SMC estimator of the normalizing constant remains unbiased, and the ARS-CSMC transition kernel leaves the target smoothing distribution invariant [2606.25603].

A more stringent diagnostic is the $\infty$-ESS,
$$
\frac{\|w\|_1}{\|w\|_\infty}
=
\frac{\sum_{n=1}^N w^n}{\max_n w^n},
$$
which is smaller than the usual $2$-ESS and therefore imposes stricter control on weight concentration [1503.00966]. The conditional adaptive-resampling particle filter uses a threshold $\zeta N$ and resamples when the $\infty$-ESS falls below it. This construction supports direct divergence control rather than only control of mean-squared error of test-function estimates. Under the $\infty$-ESS bound and a strong-mixing assumption, the cited results provide bounds on $\mathbb E[\hat Z_t]/Z_t$, a minorization inequality for the sampling kernel, and $O(1/N)$ control of KL and $\chi^2$ divergences between the target distribution and the law of the SMC sample [1503.00966].

The same framework is carried into Particle Gibbs. Replacing the conditional SMC kernel by its adaptive-resampling variant preserves invariance and yields a minorization condition implying uniform ergodicity of the conditional kernel; the full adaptive-resampling Particle Gibbs chain is then geometrically ergodic whenever the ideal Gibbs sampler is [1503.00966]. This suggests that adaptive resampling in SMC is not merely a computational heuristic: in these formulations it is tightly coupled to validity guarantees for approximate sampling and PMCMC.

## 4. Adaptive particle allocation in Resample-Move

For normalizing-constant estimation, adaptive resampling appears as adaptive expansion of the particle set within the Resample-Move framework. The motivating observation is that a fixed number of particles at every intermediate distribution is suboptimal because some transitions contribute disproportionately to the variance of the estimate [1604.01972]. Let
$$
v_n = \mathrm{Var}_{\pi_n}\!\left[\frac{W(x_{[n]})}{\mathbb E_{\pi_n}[W(x_{[n]})]}\right].
$$
Under independent equally weighted samples and a total budget $R_{\mathrm{tot}} = \sum R_n$, the optimal particle allocation is
$$
R_n^* =
\left[
\frac{\sqrt{v_n}}{\sum_{k=1}^{N-1}\sqrt{v_k}}
\right] R_{\mathrm{tot}},
\qquad
n=1,\dots,N-1,
$$
and the minimized variance is
$$
\mathrm{Var}_{\min}(\log \hat Z)
=
\frac{\left[\sum_{n=1}^{N-1}\sqrt{v_n}\right]^2}{R_{\mathrm{tot}}}.
$$

Because $v_n$ is not available a priori, Adaptive Resample-Move (ARM) uses the proxy
$$
\hat v_n \approx \frac{R_n}{\mathrm{ESS}_n} - 1,
$$
with $\gamma_n = \mathrm{ESS}_n / R_n$ as a per-particle quality diagnostic. If $\gamma_n < \gamma_{\mathrm{thr}}$, the particle set is expanded; if $\gamma_n \ge \gamma_{\mathrm{thr}}$, the current population is regarded as sufficient [1604.01972]. The algorithm therefore implements a greedy online version of the optimal allocation rule.

The benchmark results cited for Gaussian Process Classification and Restricted Boltzmann Machine partition-function estimation report that ARM estimates the normalizing constant with a smaller variance, using less computational resources, than either Resample-Move with a fixed number of particles or Annealed Importance Sampling; a further reported advantage is that ARM is easier to tune [1604.01972]. In the GPC experiments, the description states that ARM reaches low bias and variance an order of magnitude faster than AIS or fixed-$R$ RM. In the RBM experiments, ARM and especially ARM-anticipate match or beat AIS in variance using less than one minute of compute, whereas AIS-MAS takes more than ten minutes. The practical control variables are $R_{\mathrm{base}}$, $\gamma_{\mathrm{thr}}$, and $i_{\max}$ rather than $\beta$-grids or HMC leapfrog parameters.

## 5. Adaptive image resampling and spatially varying kernels

In image processing, adaptive resampling refers to content-aware interpolation rather than particle-population control. Deep adaptive image resampling for super-resolution uses a kernel-estimation FCN to predict spatially variant interpolation kernels and an adaptive resampling layer that applies the estimated kernels adaptively to each position in the image [1712.06463]. If $K_H(x,y)$ denotes the filter predicted at HR location $(x,y)$, the reconstruction is expressed as
$$
\hat I_H(x,y)
=
\sum_{(\Delta x,\Delta y)\in \Omega_f}
w_{\Delta x,\Delta y}(x,y)\;
I_L\bigl(\lfloor x/s\rfloor+\Delta x,\;\lfloor y/s\rfloor+\Delta y\bigr).
$$
The model is trained end to end with an $L_1$ reconstruction loss. A recursive extension is reported to add approximately $0.05$ dB on Set5/Set14 per extra recursion, and the method is also extended to joint image filtering [1712.06463].

A later downscaling formulation, ADK-Net, learns a unique, spatially varying resampling kernel for each output pixel and each color channel, followed by per-pixel convolution on the HR image [2511.01620]. The architecture contains a ResNet-based feature extractor and three independent channel-specific kernel generators, with kernel normalization by min-max scaling to $[0,1]$ and sum-to-one normalization. On the RealSR test set for $4\times$ downscaling, the reported results are Bicubic at $25.79$ dB / $0.8168$, DPID at $30.30$ dB / $0.8698$, IDCL at $28.10$ dB / $0.8759$, SDFlow at $32.55$ dB / $0.9466$, and ADK-Net at $33.27$ dB / $0.9560$; the ablation summary reports that sharing the trunk or using a single stream for all channels drops PSNR by more than $0.8$ dB [2511.01620].

Remote-sensing and astronomical variants use different adaptive signals. A Cellular Automata–CNN hybrid adaptive resampling scheme for remotely sensed imagery selects among $\{\mathrm{NN}, \mathrm{BL}, \mathrm{CC}, \mathrm{KD16}\}$ according to local pixel variation and texture variation, evaluated over a Laplacian pyramid; the reported comparison states that the method consistently achieves higher correlation, lower entropy deviation, and reduced average error than classical kernels [1405.6135]. For JWST IFU spectroscopy, adaptive trace modeling fits cubic B-spline profiles along slices, oversamples detector traces before cube rectification, and switches between spline-based and linear interpolation according to the derivative of the normalized spline. The reported effect is a reduction of single-spaxel artifact amplitudes from approximately $8$–$10\%$ down to approximately $1\%$, with aperture-summed flux conservation better than $0.3\%$ rms after ARS [2603.26916].

This suggests a domain-specific shift in meaning: in imaging, “adaptive resampling” usually denotes locally varying interpolation or reconstruction operators, even when the formal mechanism is still a data-dependent redistribution of sampling weight.

## 6. Adaptive resampling in learning, optimization, and proposal adaptation

Beyond SMC and image formation, adaptive resampling is used to reallocate budget across classes, features, candidate solutions, or proposals. In imbalanced classification, Adaptive Resampling-based Training (ART) periodically recomputes a class-sampling distribution from class-wise validation F$_1$ scores. With $f_i$ the per-class F$_1$, the difficulty score is $s_i = 1 - f_i$, the normalized performance-based distribution is $w_i = s_i / \sum_j s_j$, and the training distribution is blended with the empirical prior as
$$
p_i = c\,\Pi_i + (1-c)\,w_i.
$$
The reported empirical summary states that ART improves macro F1 by an average of $2.64$ percentage points across all tested tabular datasets relative to training on the original imbalanced data, and that gains on tabular datasets are significant under paired $t$-tests and Wilcoxon tests with $p < 0.05$ [2509.00955].

In iterated sampling importance resampling, the adaptive variable is the number of proposals $N$. The method defines a cost-weighted proxy
$$
\Lambda(N) = (a+bN)\,\frac{1+\epsilon(N)}{1-\epsilon(N)},
$$
where $\epsilon(N)$ is the average holding probability, proves that the asymptotic variance $V_N(f)$ is strictly decreasing and strictly convex in $N$, and introduces a generalized transition with fractional “number of proposals” to support online adaptation [2512.00220]. The adaptive algorithm updates a reparameterized proposal count by stochastic approximation and is accompanied by a strong law of large numbers for bounded test functions [2512.00220].

Adaptive resampling of random Fourier features aims to approach the optimal frequency law
$$
p_*(\tfrac{\pi}{L}n)
=
\frac{|\hat f(\tfrac{\pi}{L}n)|}
{\sum_m |\hat f(\tfrac{\pi}{L}m)|}
$$
by repeatedly resampling frequencies according to the magnitudes of learned amplitudes $\bar \beta_n$ [2509.03151]. The main theorem summarized in the source states that, under decay and regularity assumptions and suitable choices of the cutoff $\epsilon$ and fraction $q_\epsilon$, one ARS step yields a resampled law whose generalization error approaches the optimal constant $C_{p_*}(f)$ up to a factor $1 + C''\,\mathcal O(q_\epsilon^{1/2})$ [2509.03151].

In noisy multi-objective optimization, ARS uses bootstrap estimates of the means and a probability-of-dominance criterion to decide whether a point merits additional evaluation. If $p_j$ is the maximum estimated probability that $\bar{\mathbf Y}_j$ dominates a member of the current Pareto-front estimate, then a point is resampled only when $p_j$ lies between lower and upper thresholds $\alpha_\ell$ and $\alpha_u$ [2503.21495]. The empirical summary reports that ARS is better than all other sequential NSGA-II variants in $71.0\%$ of test cases and worse in $8.7\%$, while only static one-shot NSGA-II sometimes beats ARS, notably under $\chi^2$ noise [2503.21495].

A related but geometrically different use appears in multi-view 3D detection. FrustumFormer performs adaptive instance-aware resampling by constructing instance frustums from 2D masks, predicting a BEV occupancy mask, and dedicating additional deformable-attention sampling only to selected BEV queries. The component ablation reports an improvement from $0.318$ mAP / $0.366$ NDS for the baseline to $0.337$ mAP / $0.383$ NDS for the full adaptive instance-aware resampling module, and to $0.360$ mAP / $0.463$ NDS with temporal fusion on top [2301.04467].

## 7. Trade-offs, limitations, and the distinction from adaptive rejection sampling

The central trade-off in adaptive resampling is between computational cost and approximation quality. In ESS-triggered particle filters, larger thresholds induce more frequent resampling and lower degeneracy, while smaller thresholds save work but allow more weight collapse [2606.25603]. In KLD-resampling, smaller $\varepsilon$ or smaller $\delta$ imply more conservative sample sizes [1306.3172]. In ARM, more particle expansions reduce variance of $\log \hat Z$ but increase runtime [1604.01972]. In learned image resampling, larger or more specialized kernel generators can improve fidelity but increase model complexity [2511.01620]. In noisy optimization, more aggressive resampling can waste budget on points that are already confidently dominated or have too little chance to improve [2503.21495].

Another recurrent limitation is that adaptivity depends on the adequacy of the diagnostic. ESS summarizes weight concentration but not all aspects of particle impoverishment; KLD-resampling depends on binning; ARM depends on ESS as a proxy for the unknown variance term $v_n$; bootstrap-based ARS in noisy optimization assumes global homoscedasticity when the number of observations is small [1306.3172] [1604.01972] [2503.21495]. These are not objections to adaptive resampling as such, but they locate the modeling burden in the choice of the adaptation signal.

Finally, adaptive resampling should be distinguished from adaptive rejection sampling. “Adaptive Rejection Sampling with fixed number of nodes” introduces Cheap Adaptive Rejection Sampling (CARS), where the target is a univariate log-concave density, the proposal is a piecewise-exponential envelope, and the number of support nodes is held fixed to keep proposal-sampling cost constant [1509.07985]. “Parsimonious Adaptive Rejection Sampling” likewise describes an efficient trade-off between acceptance rate and proposal complexity, with the resulting algorithm reported as faster than standard ARS [1710.04948]. These methods concern adaptive proposal construction inside rejection sampling, not adaptive triggering or allocation of resampling steps. The shared acronym is therefore a source of genuine ambiguity rather than a terminological variant of the same method.

Source: https://www.emergentmind.com/topics/adaptive-resampling-ars