Ratio-Preserving Sampler Methods

• Ratio-Preserving Sampler is a class of algorithms designed to ensure that the proportions of target components, weights, or event frequencies exactly match specified ratios.
• These methods employ tailored techniques such as pairwise updates in MCMC, thinning/chopping in SMC, and PINN-based diffusion to enforce ratio constraints.
• Empirical studies show that ratio-preserving samplers improve effective sample size, stability, and uniform coverage in applications from combinatorial tasks to model predictive control.

A ratio-preserving sampler is any sampling procedure or family of algorithms designed to ensure that the proportions (ratios) of target components, weights, or event frequencies in the samples match specified or intrinsic ratios dictated by the target distribution, structure, or design goals. The ratio preservation may be enforced for discrete combinatorial objects (e.g., composition ratios), for continuous mixture components, or within the context of importance or particle weights. Ratio preservation is central to a variety of modern algorithms in Markov chain Monte Carlo (MCMC), sequential Monte Carlo (SMC), diffusion models, and model-predictive control. Implemented correctly, such samplers can ensure accurate statistical fidelity of mixture weights, uniform or balanced exploration, or robust control of weight degeneracy and effective sample size.

1. Ratio-Preserving MCMC for Compositional Sampling

A canonical context for ratio-preserving samplers is in combinatorial composition, where the goal is to generate non-negative integer-valued vectors that satisfy strict sum and sparsity constraints. Given $D$ possible components and a positive integer $N$ representing granularity, valid compositions are $x = (x_1, \ldots, x_D) \in \mathbb{N}_0^D$ with $\sum_{i=1}^D x_i = N$, and often only a small number $k = \|x\|_0$ of these entries are nonzero. The ratio-preserving MCMC sampler introduced in "Sampler for Composition Ratio by Markov Chain Monte Carlo" formalizes the following:

• Target Distribution: Defines an energy-based probability for each valid $x$, incorporating both a prescribed sparsity prior $y_\text{sparse}(k)$ and auxiliary property-based 'goodness' measures $y_\text{prop}(x)$, yielding

$$\pi(x) = \frac{1}{Z} \exp\Bigl(-E_\text{sparse}(x) - \sum_{p>1} E_\text{prop,p}(x)\Bigr).$$

The sparsity energy is

$$E_\text{sparse}(x) = -\log \frac{y_\text{sparse}(\|x\|_0)}{\binom{D}{\|x\|_0} \binom{N-1}{N-\|x\|_0}},$$

where the denominator counts the number of ways to realize the specified nonzero pattern (Obara et al., 2019).

• Proposal Mechanism: Each proposal redistributes the sum $N$ between a pair of entries $(x_i, x_j)$ while maintaining all other coordinates, choosing the pair $(i, j)$ according to a ratio-preserving weight and proposing all admissible splits of $x_i + x_j$.
• Acceptance: Acceptance probabilities are computed as

$$a(x \rightarrow x') = \min\bigl\{1, \frac{\alpha_{ij}(x')}{\alpha_{ij}(x)}\bigr\}.$$

This preserves detailed balance and ensures the stationary distribution matches the target exactly.

• Theoretical Properties: The chain is reversible, ergodic, and, under practical sparsity regimes, enjoys rapid mixing, with an acceptance rate lower bound of 0.5 for all moves.
• Empirical Verification: Tested on combinatorial creative tasks (e.g., generation of cocktail recipes subject to multiple constraints), the sampler preserves the target sparsity and property distributions and achieves high acceptance rates (95–99%) (Obara et al., 2019).

2. Ratio-Preserving Resampling Algorithms

In sequential Monte Carlo (SMC), resampling schemes are often required to curtail weight degeneracy. Standard methods enforce equal weights, but this can introduce excessive variability if the initial weights are already balanced. The ChopThin algorithm presents a ratio-preserving resampling strategy:

• Objective: Given particle weights $\{w_i\}_{i=1}^n$, output new weights $\{w_i'\}_{i=1}^N$ such that

$\max_i w_i' / \min_j w_j' \leq \eta$

for specified ratio bound $\eta \geq 1$, while ensuring unbiasedness and total weight conservation.

• Algorithm: Particles with small weights are thinned (randomly dropped), and those with large weights are chopped (split into multiple descendants), with the number of offspring determined by a piecewise linear function $h_a^\eta(w)$ and a threshold $a$ ensuring the ratio constraint (Gandy et al., 2015).
• Theoretical Guarantees: The method guarantees a lower bound on effective sample size (ESS), with

$ESS \geq \frac{4\eta n}{(\eta+1)^2}$

for $n \gg 1$, for any ratio bound, and achieves linear-time complexity.

• Performance: Simulation studies consistently show ChopThin delivers lower mean-squared error and higher stability compared to standard resamplers, especially at moderate $\eta$ (e.g., $\eta = 3 + \sqrt{8} \approx 5.8$ for $ESS \geq 0.5n$) (Gandy et al., 2015).

3. Diffusion-Based Ratio-Preserving (Mixing Proportion-Preserving) Samplers

Diffusion-based samplers are widely used for generative modeling and density estimation, but standard score-matching approaches can fail to recover mixing ratios correctly in multimodal settings. The Diffusion-PINN Sampler (DPS) addresses this issue:

• Foundation: DPS constructs a reverse SDE whose drift is parameterized by the gradient of a log-density function $u_t(x) = \log p_t(x)$, which is itself learned via a physics-informed neural network (PINN) by solving the log-density variant of the Fokker–Planck equation.
• Log-Density PINN: The neural network $u_\theta(x, t)$ is trained (via residual minimization and suitable initial/boundary conditions) so that its time and spatial derivatives satisfy the log-density PDE throughout the $(x, t)$ domain.
• Ratio Preservation: Solving the log-density PDE (as opposed to merely the score PDE) forces the PINN to encode the exact mixture weights present in $\mu(x) = \sum_i w_i p_i(x)$ at $t=0$. Theorem 4.1 shows that the learned log-density remains close to the true log-density, ensuring that the sampled trajectories reflect the correct mixing proportions.
• Empirical Results: Across multimodal benchmarks, DPS matches the true ratios virtually exactly, with KL divergence and $L^2$ error in mixing weights being 5–100× lower than in competing samplers that do not preserve ratios (Shi et al., 2024).

4. Ratio-Preserving Trajectory Samplers in Model Predictive Control

In the context of sampling-based model-predictive control, ratio preservation corresponds to generating trajectory samples such that the induced coverage over configuration space is uniform, increasing exploration diversity. The Neural C-Uniform sampler implements this property:

• C-Uniformity Definition: For any subset $S$ of a reachable set $L_t$, the probability $P(x_t \in S) = \mu(S)/\mu(L_t)$, ensuring a uniform density over all accessible configurations.
• Neural Estimation: A neural network $p_\theta(u|x)$ is trained to output action probabilities maximizing the entropy of the subsequent state's distribution, enforced by an entropy-based uniformity ratio metric $UR_t = H(p)/H_{\textrm{unif}}$.
• Long-Horizon Ratio Preservation: Empirically, trained networks preserve $UR_t > 0.9$ at horizons much longer than those seen during training, indicating robust maintenance of trajectory-level ratio preservation.
• Integration in CU-MPPI: The ratio-preserving sampler is incorporated into CU-MPPI, enhancing performance particularly in high-curvature and obstacle-rich scenarios by increasing the diversity of sampled trajectories (Poyrazoglu et al., 4 Mar 2025).

5. Comparative Properties and Algorithmic Principles

Algorithm / Domain	Preserved Ratio	Enforcement Mechanism
MCMC Ratio Sampler	Component counts	Pairwise updates, detailed balance, sparsity control
ChopThin Resampler	Particle weight ratios	Thinning/chopping with ratio-bound
Diffusion-PINN Sampler	Mixture proportions	Log-density FPE PINN, initial condition encoding
Neural C-Uniform	Trajectory cell mass	Entropy maximization, neural policies

Each algorithmic family employs a principled enforcement of ratio preservation matched to the structure and constraints of the target domain, ensuring both theoretical and empirical fidelity.

6. Implementation Notes and Empirical Outcomes

Ratio-preserving samplers are typically designed to be computationally efficient. For example, ChopThin runs in linear time via two systematic resampling passes (Gandy et al., 2015); the combinatorial MCMC sampler only recalculates local energies, accelerating updates (Obara et al., 2019). Neural ratio-preserving samplers rely on scalable policy architectures and unsupervised objectives (Poyrazoglu et al., 4 Mar 2025).

Across tested domains, empirical studies consistently show that ratio-preserving mechanisms enhance statistical consistency, exploration diversity, or estimation accuracy. In model-predictive control, uniformity of configuration space coverage is preserved even at horizons beyond the training set (Poyrazoglu et al., 4 Mar 2025). In multimodal sampling, the proportion of samples in each component matches the desired mixture weights (Shi et al., 2024), correcting issues inherent to score-only samplers.

7. Theoretical Guarantees and Practical Recommendations

Ratio-preserving samplers are formally justified via reversibility, ergodicity, or PDE-based approximation bounds. For SMC, selecting the ratio bound $\eta$ directly trades off between effective sample size and variability, with explicit formulas guiding parameter choices (Gandy et al., 2015). In diffusion-based settings, bounds on the residual loss in the PINN directly translate to explicit convergence rates in total sample quality and mixing ratio fidelity (Shi et al., 2024).

When implementing such algorithms, practitioners are advised to:

• Exploit local structure for efficient updates (e.g., pairwise changes in MCMC (Obara et al., 2019)).
• Parameterize ratio bounds based on ESS or application-driven loss functions (Gandy et al., 2015).
• Use neural ratio-preserving samplers to mitigate the curse of dimensionality present in flow-based trajectory uniformizers (Poyrazoglu et al., 4 Mar 2025).
• Leverage problem-specific regularization and stopping criteria in PINN-based samplers to ensure coverage of the target’s support (Shi et al., 2024).

The ratio-preserving sampler family thus constitutes a unifying framework across combinatorial, continuous, and sequential domains, enabling precise control over mixture weights, component proportions, and spatial uniformity in both statistical and algorithmic contexts.