Parameterized Sampling Methods

Updated 14 July 2026

Parameterized sampling is a methodological framework where the sampling process is controlled through explicit finite-dimensional parameters, enabling adaptable and optimizable sampling strategies.
It encompasses a range of techniques—from differentiable subset selection and Bayesian posterior sampling to adaptive proposal designs and transport-based density evolution—each tailored to specific applications.
Practical implementations demonstrate measurable improvements in computational efficiency and model performance across various fields such as machine learning, inverse problems, and control.

Searching arXiv for recent and relevant papers on “parameterized sampling” across its major usages. arXiv search results (selected):

(Dua et al., 2 Oct 2025) — “FlexDoc: Parameterized Sampling for Diverse Multilingual Synthetic Documents for Training Document Understanding Models”
(Xie et al., 2019) — “Reparameterizable Subset Sampling via Continuous Relaxations”
(Gopalan et al., 2014) — “Thompson Sampling for Learning Parameterized Markov Decision Processes”
(Abbasi-Yadkori et al., 2014) — “Bayesian Optimal Control of Smoothly Parameterized Systems: The Lazy Posterior Sampling Algorithm”
(Morzfeld et al., 2013) — “Parameter estimation by implicit sampling”
(Morzfeld et al., 2016) — “Iterative importance sampling algorithms for parameter estimation”
(Wu et al., 2023) — “Parameterized Wasserstein Hamiltonian Flow”
(Correnty et al., 2023) — “Chebyshev HOPGD with sparse grid sampling for parameterized linear systems”
(Esmer et al., 2024) — “Sampling with a Black Box: Faster Parameterized Approximation Algorithms for Vertex Deletion Problems”
(Haug et al., 2021) — “Natural parameterized quantum circuit”
(Muniraju et al., 2018) — “Coverage-Based Designs Improve Sample Mining and Hyper-Parameter Optimization”
(Chen et al., 21 Jun 2026) — “Extended parameterized spin expansion formalism for ringdown analysis with GW250114” Parameterized sampling denotes a family of methods in which either the sampling distribution, the sampled combinatorial object, or the generative mechanism is controlled by explicit parameters. In the literature, the expression covers several distinct but structurally related practices: reparameterizable subset selection from weighted distributions, posterior sampling over latent model parameters, importance sampling with parameterized proposal families, push-forward sampling by parametric transport maps, and application-specific generators whose outputs are steered by schema, geometry, or physical parameters (Xie et al., 2019, Gopalan et al., 2014, Morzfeld et al., 2013, Wu et al., 2023, Dua et al., 2 Oct 2025). This suggests that “parameterized sampling” is not a single formalism, but a recurring methodological pattern in which sampling is made controllable, optimizable, or analyzable through a finite-dimensional parameterization.

1. Scope and unifying structure

Across the cited literature, parameterized sampling appears in several non-equivalent senses.

Usage in the literature	Parameterized object	Representative papers
Differentiable discrete sampling	Weighted subset distributions and relaxed $k$ -hot selections	(Xie et al., 2019)
Bayesian posterior sampling	Unknown MDP or control parameters sampled from posteriors	(Gopalan et al., 2014, Abbasi-Yadkori et al., 2014, Gimelfarb et al., 2023)
Importance and transport sampling	Proposal families $q(\theta)$ , implicit maps, or push-forward maps $T_\theta$	(Morzfeld et al., 2013, Morzfeld et al., 2016, Wu et al., 2023)
Structured design and synthesis	Coverage-based point sets, sparse-grid snapshot sets, synthetic documents	(Muniraju et al., 2018, Correnty et al., 2023, Dua et al., 2 Oct 2025)

A common motif is that a difficult target object is not sampled directly. Instead, one introduces a parametric family—weights $\mathbf{w}$ , posterior $\pi_t(\theta)$ , proposal parameters $(\mu,\Sigma)$ , transport parameters $\theta$ , or schema controls—and sampling is performed through that family. In some cases the parameters are optimized end-to-end; in others they are updated by Bayes’ rule, by moment matching, or by geometric constraints.

A frequent misconception is to equate parameterized sampling with the reparameterization trick alone. The cited work does not support that reduction. Reparameterizable subset sampling is one important instance, but the term is also used for Thompson-style posterior sampling in reinforcement learning, implicit sampling for Bayesian inverse problems, parameterized transport dynamics on probability manifolds, and several domain-specific sampling constructions (Xie et al., 2019, Gopalan et al., 2014, Morzfeld et al., 2013, Wu et al., 2023).

2. Differentiable subset sampling and continuous relaxations

In machine learning, a particularly precise meaning of parameterized sampling is sampling a discrete subset from a distribution whose weights are outputs of a differentiable model. In "Reparameterizable Subset Sampling via Continuous Relaxations" (Xie et al., 2019), a subset of size $k$ is represented as a $k$ -hot vector $S \in \{0,1\}^n$ with $q(\theta)$ 0, and the distribution is parameterized by nonnegative weights $q(\theta)$ 1. The core difficulty is that subset sampling without replacement is discrete and combinatorial, so direct stochastic optimization is not reparameterizable.

The construction generalizes Gumbel-max from a single category to a top- $q(\theta)$ 2 mechanism. Independent Gumbel perturbations are added to $q(\theta)$ 3, and a differentiable RelaxedTopK operator is applied to the perturbed scores, yielding a relaxed $q(\theta)$ 4-hot vector $q(\theta)$ 5 with $q(\theta)$ 6. As temperature tends to zero, the relaxed samples converge in distribution to exact samples from the weighted sampling-without-replacement distribution. The method also proves an order-preservation property: for the specified top- $q(\theta)$ 7 relaxation, if $q(\theta)$ 8, then $q(\theta)$ 9 for temperatures $T_\theta$ 0. The paper therefore extends reparameterization from categorical variables to structured subset variables.

The same work demonstrates that the construction is not merely formal. In instance-wise feature selection on IMDB, RelaxSubSample improves post-hoc accuracy from $T_\theta$ 1 to $T_\theta$ 2 on IMDB-word and from $T_\theta$ 3 to $T_\theta$ 4 on IMDB-sent. In deep stochastic $T_\theta$ 5-nearest neighbors, it reaches test accuracies of 99.3 on MNIST, 93.6 on Fashion-MNIST, and 90.1 on CIFAR-10, while remaining markedly faster than NeuralSort at large candidate-set size; for example, at $T_\theta$ 6, forward-pass runtime is 0.110s versus 3.694s. In RSS-SNE, trustworthiness $T_\theta$ 7 on MNIST reaches 0.947 at $T_\theta$ 8 and 0.999 at $T_\theta$ 9, compared with 0.926 and 0.983 for parametric t-SNE, while 1-NN error drops from 9.90% to 8.31% at $\mathbf{w}$ 0 and from 5.41% to 3.05% at $\mathbf{w}$ 1. Parameterized subset sampling is therefore both a geometric relaxation and a computational primitive for end-to-end learning.

3. Posterior sampling in parameterized decision processes

In reinforcement learning and control, parameterized sampling often means sampling an unknown environment parameter from a posterior and acting optimally for the sampled model. In "Thompson Sampling for Learning Parameterized Markov Decision Processes" (Gopalan et al., 2014), the environment is a family of MDPs $\mathbf{w}$ 2, where the parameterization induces correlations across transitions and rewards. Thompson sampling proceeds by sampling $\mathbf{w}$ 3 from the posterior at each epoch, computing an average-reward optimal policy $\mathbf{w}$ 4, and executing it until the next return to a reference state. The principal theoretical result is that the number of suboptimal actions grows logarithmically with time, with high probability, and the constant factor is determined by an optimization problem involving marginal KL divergences over the policy class. The paper also derives an average-reward regret bound and emphasizes that the relevant complexity is the KL geometry of the parameter space rather than tabular state-action cardinality.

"Bayesian Optimal Control of Smoothly Parameterized Systems: The Lazy Posterior Sampling Algorithm" (Abbasi-Yadkori et al., 2014) adopts the same posterior-sampling logic in a more general controlled dynamical setting, but makes computation central. The unknown dynamics are parameterized by $\mathbf{w}$ 5, and the algorithm samples $\mathbf{w}$ 6 from the posterior only when the determinant of an information matrix doubles. The resulting Lazy PSRL policy updates only $\mathbf{w}$ 7 times, while the regret is $\mathbf{w}$ 8, where $\mathbf{w}$ 9 is cumulative planning suboptimality. Here parameterized sampling is not just an exploration device; it is a mechanism for trading statistical efficiency against repeated solution of expensive control problems.

A later variant, "Thompson Sampling for Parameterized Markov Decision Processes with Uninformative Actions" (Gimelfarb et al., 2023), studies PMDPs in which some actions yield no information about the unknown parameter. The paper identifies assumptions under which Thompson sampling still guarantees asymptotically optimal expected regret of $\pi_t(\theta)$ 0. Its posterior concentration result is exponential: $\pi_t(\theta)$ 1. The key structural condition is the existence of a special state $\pi_t(\theta)$ 2 such that the optimal action at $\pi_t(\theta)$ 3 is informative for every parameter, together with recurrence assumptions ensuring that such informative state-action pairs are visited often enough. Queueing, inventory control, and dynamic pricing are treated as exemplars. This use of parameterized sampling is therefore Bayesian, model-based, and policy-mediated.

4. Parameterized proposals in Bayesian inverse problems

In inverse problems, parameterized sampling typically refers to proposal distributions or transformations whose finite-dimensional parameters are adapted to posterior geometry. "Parameter estimation by implicit sampling" (Morzfeld et al., 2013) formulates the posterior as $\pi_t(\theta)$ 4 and defines $\pi_t(\theta)$ 5. Rather than constructing a Markov chain, implicit sampling first finds the MAP point $\pi_t(\theta)$ 6, then maps independent reference samples $\pi_t(\theta)$ 7 to parameter samples $\pi_t(\theta)$ 8 through the implicit relation $\pi_t(\theta)$ 9. Two concrete maps are given: a linear map based on a local quadratic model $(\mu,\Sigma)$ 0, and a random map $(\mu,\Sigma)$ 1 whose Jacobian enters the importance weights. The method is then instantiated for an elliptic PDE inverse problem with a KL parameterization of the log-permeability field; the truncation dimension is $(\mu,\Sigma)$ 2, capturing 99.9% of the variance. With $(\mu,\Sigma)$ 3 samples, the linear map yields $(\mu,\Sigma)$ 4 and the random map $(\mu,\Sigma)$ 5, where effective sample size is approximately $(\mu,\Sigma)$ 6. The paper also reports that stable posterior mean estimates are obtained with about $(\mu,\Sigma)$ 7 implicit samples and roughly $(\mu,\Sigma)$ 8 forward solves, whereas a Metropolis chain had not stabilized even after $(\mu,\Sigma)$ 9 steps.

"Iterative importance sampling algorithms for parameter estimation" (Morzfeld et al., 2016) makes the parameterization of the proposal distribution explicit. The proposal is either multivariate Gaussian $\theta$ 0 or multivariate $\theta$ 1, and its parameters are updated by weighted moment matching. Proposal quality is measured by $\theta$ 2, with effective sample size $\theta$ 3. The method can be initialized by short MCMC runs or by Gaussian mixture models around local optima. On a 14-dimensional subsurface-flow problem, $\theta$ 4 decreases from 163 to 1.3 in 9 iterations using 46,080 samples, with wall-clock time about 90 minutes; the comparison MCMC run required about 20 days. On a 9-parameter combustion problem, a $\theta$ 5-proposal with $\theta$ 6 reduces $\theta$ 7 from 293 to 27 after a few iterations, and each iteration with $\theta$ 8 samples takes about 15 minutes on 3,072 cores. In this setting, parameterized sampling is an adaptive importance-sampling architecture in which the proposal family itself is the object of inference.

5. Push-forward dynamics and reduced-order sampling over parameter spaces

A different use of parameterized sampling appears when a probability measure or a PDE solution manifold is represented by a finite-dimensional map and sampling occurs through that map. "Parameterized Wasserstein Hamiltonian Flow" (Wu et al., 2023) defines a finite-dimensional manifold of densities by $\theta$ 9, where $k$ 0 is a parameterized push-forward map and $k$ 1 is a reference measure. Pulling the Wasserstein Hamiltonian structure back to parameter space yields a Hamiltonian $k$ 2, together with a finite-dimensional Hamiltonian ODE. For numerical work, the paper uses a simplified pullback metric $k$ 3, solves the resulting ODE by a symplectic Euler scheme, and emphasizes that the method is fully deterministic and involves no neural-network training or SGD. Sampling from the approximate density is immediate: draw $k$ 4, then output $k$ 5. The method comes with Wasserstein error bounds involving expressivity quantities $k$ 6, and it is demonstrated on examples including a 10-dimensional harmonic oscillator. Parameterized sampling here means that the sampler is the time-evolved transport map.

"Chebyshev HOPGD with sparse grid sampling for parameterized linear systems" (Correnty et al., 2023) addresses parameterized linear systems $k$ 7 by constructing sparse-grid snapshot sets and low-rank tensor surrogates. The sampling grid is cross-shaped: $k$ 8 and $k$ 9. Companion linearization and preconditioned BiCG generate many snapshots along one parameter line from a single Krylov run, and the snapshot tensor is decomposed as

$k$ 0

Spline interpolation of the scalar factors yields a reduced-order model that is cheap to evaluate over the full parameter domain. In the Helmholtz example, a ROM evaluation takes 0.0106 CPU seconds, compared with 1.1979 seconds for a full finite-element solve, and in the denser-sampling experiment all 400 test points have relative error below 1%. This is parameterized sampling in the sense of strategically sampling a parameter space to construct a separable approximation of a whole solution family.

6. Design, synthesis, and combinatorial search

In some areas, parameterized sampling means designing the sample set itself through a parametric model of coverage or success probability. "Coverage-Based Designs Improve Sample Mining and Hyper-Parameter Optimization" (Muniraju et al., 2018) introduces a parameterized family of coverage-based designs through a pair correlation function

$k$ 1

combining a hard-core radius $k$ 2, a peak between $k$ 3 and $k$ 4, and damped oscillations beyond $k$ 5. The associated power spectral density is constrained to be nonnegative, and the parameters are optimized to maximize coverage radius under realizability constraints. A gradient-based synthesis algorithm with adaptive learning rate then matches an empirical PCF to the target. The resulting designs achieve a 25–40% increase in coverage radius over conventional Poisson disk sampling and outperform discrepancy-based exploratory designs in both blind exploration and Bayesian optimization for sample mining and hyper-parameter tuning.

A more explicitly algorithmic use appears in "Sampling with a Black Box: Faster Parameterized Approximation Algorithms for Vertex Deletion Problems" (Esmer et al., 2024). There, a sampling step is a polynomial-time randomized routine returning a vertex $k$ 6 such that $k$ 7 drops by 1 with probability at least $k$ 8. Combined with either an exact FPT black box or a polynomial-time $k$ 9-approximation, repeated sampling yields parameterized $S \in \{0,1\}^n$ 0-approximation algorithms whose exponents are optimized through a KL-divergence analysis. For Feedback Vertex Set, the paper shows a sampling step with $S \in \{0,1\}^n$ 1 and derives, for every $S \in \{0,1\}^n$ 2, a faster parameterized $S \in \{0,1\}^n$ 3-approximation than the MFCS 2023 result of Jana, Lokshtanov, Mandal, Rai, and Saurabh. The same framework also treats $S \in \{0,1\}^n$ 4-Hitting Set and $S \in \{0,1\}^n$ 5-Path Vertex Cover.

A domain-specific generative use is given by "FlexDoc: Parameterized Sampling for Diverse Multilingual Synthetic Documents for Training Document Understanding Models" (Dua et al., 2 Oct 2025). FlexDoc combines Stochastic Schemas and Parameterized Sampling to generate realistic, multilingual semi-structured documents with rich annotations by probabilistically modeling layout patterns, visual structure, and content variability. On KIE tasks, the paper reports absolute F1 improvements of up to 11% when synthetic data augment real datasets, and annotation effort is reduced by over 90% compared to traditional hard-template methods. Here the sampled object is not a subset or a posterior parameter, but a document instance drawn from a controllable synthetic document family.

7. Physical, spectral, and domain-specific extensions

In quantum information, "Natural parameterized quantum circuit" (Haug et al., 2021) defines a parameterized quantum circuit whose quantum Fisher information matrix satisfies $S \in \{0,1\}^n$ 6 at a reference parameter. At that point, the ordinary gradient equals the quantum natural gradient, and the circuit achieves the minimal quantum Cramér–Rao bound within a broad class of Pauli-rotation circuits. The paper also develops a sampling-based parameter estimation scheme: after a modified NPQC and its adjoint are applied, computational-basis sampling yields probabilities $S \in \{0,1\}^n$ 7, giving the estimator $S \in \{0,1\}^n$ 8. This turns parameterized sampling into a calibration and benchmarking protocol for NISQ hardware.

In gravitational-wave ringdown analysis, "Extended parameterized spin expansion formalism for ringdown analysis with GW250114" (Chen et al., 21 Jun 2026) extends ParSpec by jointly sampling an effective length scale $S \in \{0,1\}^n$ 9 and scaling index $q(\theta)$ 00, with coupling $q(\theta)$ 01. The paper finds that the posterior of $q(\theta)$ 02 remains largely prior dominated, while constraints on $q(\theta)$ 03 are shaped mainly by the geometry of the allowed parameter space induced by the $q(\theta)$ 04 condition. Under $q(\theta)$ 05, the most natural prescription in the $q(\theta)$ 06 plane, the upper bound is weak but stable, $q(\theta)$ 07. A KL-divergence analysis further shows that, after accounting for the effective prior associated with $q(\theta)$ 08, the data-driven information gain on $q(\theta)$ 09 and $q(\theta)$ 10 is very small. This use of parameterized sampling is therefore a cautionary example: apparent constraints may be dominated by prior geometry rather than likelihood information.

A mathematically different but conceptually related construction is "Sampling expansions associated with quaternion difference equations" (Cheng et al., 2019). Starting from a quaternion difference equation with boundary conditions, the paper builds a parameterized sequence $q(\theta)$ 11 in $q(\theta)$ 12. For fixed $q(\theta)$ 13, there exist $q(\theta)$ 14 distinct quaternions $q(\theta)$ 15 such that $q(\theta)$ 16 forms an orthogonal basis of $q(\theta)$ 17, and every transform $q(\theta)$ 18 admits an exact sampling expansion

$q(\theta)$ 19

The sample nodes are zeros of a quaternion polynomial associated with the boundary-value problem and, equivalently, eigenvalues of a tridiagonal quaternion matrix. In this setting, parameterized sampling refers to spectral reconstruction from a parameterized kernel sequence rather than stochastic generation.

Taken together, these uses show that the term is intrinsically plural. It can denote differentiable relaxations of discrete draws, Bayesian posterior sampling, adaptive proposal design, transport-based density evolution, geometry-controlled sample design, randomized search over parameterized solution spaces, or parameterized spectral reconstruction. A plausible implication is that the unifying content of parameterized sampling lies not in a shared algorithmic template, but in a shared commitment to exposing the latent control variables of a sampling process so that they can be optimized, inferred, or constrained.