PARIS Algorithm: Adaptive Importance Sampling

• PARIS algorithm is an adaptive importance sampling method that constructs Gaussian mixture proposals using previously weighted samples for efficient posterior approximation.
• It leverages parallel ARIS processes seeded via Latin hypercube sampling and sliding-window reweighting to reduce computational cost and ensure scalability.
• Empirical results in gravitational-wave parameter estimation and high-dimensional Gaussian mixtures demonstrate its superior performance compared to traditional inference methods.

PARIS Algorithm

The umbrella term "PARIS algorithm" encompasses a set of rigorously defined methodologies spanning diverse fields, notably Bayesian sampling, imbalanced regression, large-scale data series indexing, and more. The following article focuses in detail on the most recent major algorithmic innovation: Parallel Adaptive Reweighting Importance Sampling (PARIS) (Liu et al., 22 Mar 2026), while distinguishing it from other PARIS-named methods as appropriate.

1. Conceptual Foundations and Motivation

Parallel Adaptive Reweighting Importance Sampling (PARIS) is a principled, adaptive importance sampling (IS) methodology targeting the efficient construction of approximate posterior distributions and evidence estimates for complex, high-dimensional, and frequently multi-modal targets. Standard IS methods require a proposal density $q(x)$ that closely approximates the target $P(x)$, which is often infeasible for highly multimodal or correlated spaces. Adaptive IS families (such as AMIS and population Monte Carlo) attempt to alleviate this by iteratively refining $q$ using previous samples, but they typically impose parametric structure (e.g., Gaussian mixtures with fixed components) and incur substantial computational cost when retrospectively reweighting all samples under the evolving proposal sequence, scaling as $O(T^2)$ for $T$ total samples.

PARIS avoids both of these limitations via an architecture that:

• Models each proposal $q_t(x)$, at iteration $t$, as a Gaussian mixture centered at the previous samples, with mixture weights proportional to the latest normalized importance weights.
• Leverages a recursive, deterministic-mixture reweighting scheme ensuring convergence to the target despite the adaptivity.
• Enables scalable parallelism by running multiple ARIS (Adaptive Reweighting IS) processes independently, seeding them via low-discrepancy methods and merging them as their proposals begin to overlap.
• Achieves near-linear per-iteration cost using efficient sliding-window bookkeeping and covariance adaptation.

This leads to an effective, massively parallel inference engine suitable for high-dimensional, complex inference problems such as gravitational-wave parameter estimation.

2. Core Methodology and Mathematical Formalism

The PARIS procedure is defined by the following components and update equations:

Proposal Construction

At iteration $t$ (of a given ARIS process), the proposal density is a sample-centric Gaussian mixture: $$q_t(x) = \sum_{i=1}^{t-1} w_i^{(t-1)}\ \mathcal N(x \mid x_i, \Sigma_{t-1})$$

• Mixture centers: the $t-1$ previous samples $P(x)$0.
• Weights: latest normalized importance weights $P(x)$1.
• Covariance: empirical weighted sample covariance

$P(x)$2

Covariance updates are performed every $P(x)$3 iterations, together with boundary regularization and shrinkage for stability in high dimensions.

Parallel and Sliding-Window Operations

• Multiple ARIS processes are seeded from high-posterior regions found using Latin hypercube sampling, each process maintaining its own local pool of recent samples (window size $P(x)$4).
• After each iteration, processes examine mutual proximity (in sample/proposal density) and merge if they begin to cover the same mode.

Deterministic-Mixture Reweighting

The importance weights for all live samples (across past and present proposals) at iteration $P(x)$5 are updated as: $P(x)$6 Efficiently, a recursive formula applies: $P(x)$7 All weights are re-normalized after update.

Merging Logic and Redundancy Elimination

At each outer cycle, the density $P(x)$8 of every process $P(x)$9 is evaluated at the location of all other active process's latest samples, $q$0. If $q$1, process $q$2 is said to "cover" $q$3, and only the process with maximal posterior is retained within a merged cluster. This mechanism eliminates wasted computational effort on modes already well-explored.

Posterior and Evidence Estimation

After $q$4 total iterations (summed over all ARIS processes), the posterior is estimated as the normalized set of weighted samples, and the marginal likelihood (evidence) as

$q$5

3. Detailed Algorithmic Workflow

A practical implementation of PARIS may be summarized as follows:

$O(T^2)$8 All numerical criticalities (e.g., boundary effects, windowing, shrinkage, density evaluations) follow explicit details from (Liu et al., 22 Mar 2026).

4. Computational Complexity and Optimizations

Naive adaptive methods—such as AMIS—require $q$6 storage and $q$7 total density evaluations, due to necessitating full reweighting of all prior samples against every prior proposal. PARIS reduces both cost and memory by:

• Restricting the mixture in $q$8 to a sliding window of most recent $q$9 samples.
• Using analytic weights for only those samples within each window, and efficient rolling sum updates.
• Scaling the per-iteration cost at $O(T^2)$0 (per process), with process merging scaling as $O(T^2)$1, far superior for large $O(T^2)$2.
• Normalization and covariance updates are carried out periodically and in-place.

Self-term regularization of mixture components ($O(T^2)$3 factor applied to the central peak) circumvents overweighting new modes in high-$O(T^2)$4.

5. Handling Multimodal and High-Dimensional Distributions

PARIS is explicitly constructed for, and empirically validated on, highly multimodal and strongly correlated targets. Proposals automatically concentrate mixture mass at high-posterior regions, while samples in sparse regions that nevertheless acquire high weights are promoted to mixture centers in subsequent iterations. This self-correcting mechanism ensures both aggressive exploration and robust exploitation. Parallel ARIS processes ensure full coverage of separated modes; the merge phase guarantees computational effort is not duplicated.

Empirical results across a battery of synthetic (e.g., disconnected “NUS” support, high-dimensional Gaussian mixtures) and realistic (LIGO/Virgo GW150914, galactic binary searches) benchmarks show order-of-magnitude reductions in function evaluations required to achieve state-of-the-art posterior accuracy and evidence estimation, compared to dynamic nested sampling or parallel-tempered MCMC.

Performance Table (from (Liu et al., 22 Mar 2026)):

Problem	Function Evals (PARIS)	DNS	PTMCMC	log Z (exact/PARIS)
2D NUS	10,100	1,500,000	87,000	-2.40 / -2.45
10D GMM (10 modes)	160,050	8.6M	820,000	2.30 / 2.30
GW150914 (13D)	100,000	3.1M	3.2M	-32,108.4 (both)

6. Theoretical Guarantees and Practical Recommendations

The deterministic-mixture reweighting guarantees convergence to the true target under mild regularity provided the union of proposals covers the support of $O(T^2)$5. Self-correcting weights ensure initial bias in low-sampled regions is eliminated as more samples are drawn, and the effective sample size remains high until the full mass is accounted for.

Practical usage requires:

• Initial seeding with diverse, high-quality samples (typically from Latin hypercube design with initial posterior evaluation).
• Sensible window size $O(T^2)$6 (sufficiently large to accommodate intra-mode exploration, yet small enough to ensure efficiency).
• Covariance shrinkage and boundary adaptation to preserve stability at high dimension.

It is essential to monitor merging dynamics and evidence stabilization as diagnostic of convergence.

7. Relation to Other Algorithms Named PARIS

The name "PARIS algorithm" is used for methods in:

• Pruning for imbalanced regression via representer theorem (Camporeale, 7 Dec 2025): An exact, rank-one Cholesky downdating-based iterative data pruning scheme optimizing rare-event regression generalization.
• Parallel data-series indexing (Peng et al., 2020): A SIMD/multicore disk-based time-series index construction and query answering method.
• Probabilistic alignment of relations/instances/schema (Suchanek et al., 2011): A parameter-free, scalable, holistically probabilistic ontology aligner.
• Personalized activity recommendation for sleep (Singh et al., 2021): Unsupervised timeseries clustering linked to personal outcome-driven recommendations.
• GPU inference-server partitioning (Kim et al., 2022): An integer-program inspired fast heuristic for partition scheduling on reconfigurable GPUs.
• Particle-based rapid incremental smoothing (Olsson et al., 2014): An online, $O(T^2)$7 sequential Monte Carlo smoother for additive state functionals in HMMs with provable variance bounds.

Each "PARIS algorithm" in these contexts is technically unrelated; the unifying feature is rigorous exploitation of problem structure—whether in proposal adaptation, data pruning, symbolic matching, or parallelization strategy—and the use of theoretically-grounded, efficient computations tailored to the challenges of the domain.