Realistic Distribution Setting (RDS)

• Realistic Distribution Setting (RDS) is a design principle that structures synthetic data to mimic true real-world generation processes with domain-specific constraints.
• It preserves key characteristics such as multivariate dependencies, temporal order, and realistic perturbations, as illustrated in RCT, EEG, and power network applications.
• The framework improves model evaluation and robustness by aligning synthetic data generation with authentic distributional properties observed in practical deployments.

Searching arXiv for papers using or discussing “Realistic Distribution Setting (RDS)” and closely related terminology across domains. In the recent arXiv literature, the term or idea of a realistic distribution setting is used for settings in which synthetic data, benchmark data, or target domains are constructed to preserve real-world distributional structure rather than idealized abstractions. This suggests a common usage: realism is tied to fidelity to the observed multivariate distribution, the temporal or causal order of data generation, domain-specific constraints, and deployment-relevant perturbations, rather than to row-wise plausibility or a small set of abstract graph statistics (Petrakos et al., 29 Jan 2025, Wagh et al., 2022, Meyur et al., 2022, Meyur et al., 2020).

1. Conceptual scope

A realistic distribution setting is motivated by the observation that many evaluation and synthesis pipelines become misleading when they ignore how data are actually generated or how deployment conditions actually vary. In randomized controlled trials, the relevant structure includes baseline covariates, random treatment assignment, post-randomization trajectories, and outcomes generated conditionally on earlier variables. In clinical EEG, the relevant structure is not arbitrary corruption but instrumentation-related variability in acquisition conditions. In synthetic power distribution networks, the relevant structure is geographic embedding, engineering feasibility, and economic plausibility (Petrakos et al., 29 Jan 2025, Wagh et al., 2022, Meyur et al., 2022).

In this sense, realism is distributional rather than merely visual or predictive. The RCT literature explicitly defines “realistic synthetic tabular data” as synthetic observations that reproduce the real data’s distributional features at both the univariate and multivariate levels. For RCTs, this includes preserving baseline distributions, random treatment allocation, longitudinal post-treatment trajectories, and outcome relationships while avoiding impossible values and spurious dependencies. The EEG literature formulates the problem as covariate shift with unchanged concept mapping,

$$p_s(y \mid x)=p_t(y \mid x), \qquad p_s(x)\neq p_t(x),$$

and operationalizes the target domain by synthetic but domain-guided transforms,

$p_t(x)=p_s(\mathcal{T}(x)).$

The power-systems literature defines realism through geographically embedded distribution-network datasets synthesized from real infrastructure and constrained by engineering and economic principles (Petrakos et al., 29 Jan 2025, Wagh et al., 2022, Meyur et al., 2022).

2. Structural principles of realism

A recurring principle is that the generation or evaluation pipeline should follow the same ordering as the underlying process. In the RCT setting, baseline variables are generated jointly first, treatment is generated next as a randomized allocation mechanism independent of baseline covariates, and post-treatment variables and outcomes are then generated conditional on earlier variables. The paper argues that simultaneous generation can leak information from future variables into past ones, which is antithetical to an RCT and may distort treatment assignment or baseline distributions (Petrakos et al., 29 Jan 2025).

For the baseline block, the copula formulation is used precisely because it separates marginals from dependence. The paper recalls Sklar’s theorem,

$$F_{X_1,\dots,X_d}(x_1,\dots,x_d)=C(F_{X_1}(x_1),\dots,F_{X_d}(x_d)),$$

with joint density decomposition

$$f_{X_1,\dots,X_d}(x_1,\dots,x_d)=c(F_{X_1}(x_1),\dots,F_{X_d}(x_d))\prod_{i=1}^d f_{X_i}(x_i).$$

The practical logic is then to sample from uniforms, transform through the fitted copula to dependent uniforms, and map back through inverse empirical marginals to the original scale. The copula is intended to match the joint baseline distribution while leaving subsequent treatment and outcome generation to a trial-aware mechanism (Petrakos et al., 29 Jan 2025).

In EEG, the corresponding principle is that robustness assessment should target plausible deployment-time variation in clinical acquisition rather than generic synthetic corruptions. The selected transformations are hardware band-pass filter shift, quantization precision, narrow-band impedance noise, and broadband noise. The paper explicitly contrasts these with standard augmentations such as amplitude scaling, time shifts, DC shifts, or band-stop filtering, which are considered too simplistic because they do not correspond well to instrumentation and acquisition-specific variation in clinical EEG (Wagh et al., 2022).

In synthetic power distribution networks, the same structural principle appears as bottom-up synthesis from interdependent infrastructures. Residential building locations are treated as end customers, OpenStreetMap roads are used as a spatial proxy for where distribution lines are likely to run, and high-voltage substations are used as roots of the medium-voltage network. The resulting system is required to be tree-structured, road-aligned, and voltage-feasible, rather than merely statistically similar to benchmark feeders (Meyur et al., 2022, Meyur et al., 2020).

3. Major research instantiations

Three arXiv lines make the idea concrete in different ways: synthetic tabular RCT data, realistic EEG shift evaluation, and synthetic power distribution grids (Petrakos et al., 29 Jan 2025, Wagh et al., 2022, Meyur et al., 2022, Meyur et al., 2020).

Domain	Realism mechanism	Representative result
Randomized controlled trials	Sequential generation with R-vine copula baseline, random treatment, regression-based post-treatment variables	R-vine copula sequential framework was best overall
Clinical EEG	Domain-guided target shifts from acquisition-related transforms	Stronger shifts reduced latent-space integrity and tracked performance degradation
Power distribution networks	Open geographic data plus engineering and economic constraints, with feasible ensembles	Synthetic networks were validated against actual distribution networks

In the RCT case, seven generation frameworks were compared: one simultaneous CTGAN, five sequential CTGAN variants, and an R-vine copula sequential framework. The best-performing setup used an R-vine copula model to generate baseline variables, followed by simple random treatment allocation with equal probabilities across treatment arms, and then regression models for post-treatment variables and the binary endpoint. The authors report that this framework was strongest especially for baseline distributions and bivariate relationships, whereas the simultaneous CTGAN performed poorly especially on treatment assignment because it learned from all variables at once and therefore tended to spuriously condition treatment on variables collected after randomization (Petrakos et al., 29 Jan 2025).

In the EEG case, three encoders were evaluated: a fully supervised encoder based on ShallowNet, a self-supervised encoder also based on ShallowNet, and a traditional power-spectral-density encoder using 19 channels by 7 frequency-band features. These were tested on binary EEG grade classification and brain-age regression, trained on TUAB without shifts, evaluated under synthetic shifts on held-out TUAB, and also tested on the external NMT dataset. The paper finds that stronger shifts, especially broadband noise with $\sigma=0.1$ and the tighter $1\!-\!25$ Hz band-pass filter, reduce latent-space integrity and often track the same degradation seen in task performance (Wagh et al., 2022).

In power networks, the synthesis pipeline is explicitly modular: construct the low-voltage secondary network, construct the medium-voltage primary network, generate an ensemble of feasible alternatives, and attach electrical and phase attributes. The 2020 paper formulates primary and secondary construction as optimization problems with structural and power-flow constraints. The 2022 paper extends this by producing ensembles of feasible primary networks through a Markov chain over feasible states, so that the output is not a single deterministic graph but many plausible realizations for the same region (Meyur et al., 2022, Meyur et al., 2020).

4. Validation strategies

A realistic distribution setting requires validation criteria that directly test the preserved structure. In the RCT framework, univariate fidelity is measured by the complement of the Kolmogorov–Smirnov statistic for continuous variables and the complement of total variation distance for discrete variables. Bivariate fidelity is assessed with a correlation similarity score for continuous-continuous pairs and a contingency-table similarity score for discrete-discrete or mixed pairs. The study also reports machine-learning utility metrics—precision, recall, and F1-score using XGBoost and k-nearest neighbors—and total compute time. Across these evaluations, the R-vine sequential framework usually gave the best or most stable scores, especially for univariate distributions, bivariate dependence, and preserving the randomized treatment distribution (Petrakos et al., 29 Jan 2025).

The EEG study adds two diagnostics beyond task accuracy. For latent space integrity, it constructs a Delaunay neighborhood graph over original and transformed embeddings and computes the proportion of heterogeneous edges,

$$q(\mathcal{G}) = 1-\frac{\left(\left|\mathcal{G}^{Z}\right|_{\mathcal{E}} + \left|\mathcal{G}^{\tilde Z}\right|_{\mathcal{E}}\right)}{\left|\mathcal{G}\right|_{\mathcal{E}}},$$

with higher $q(\mathcal{G})$ interpreted as better integrity under perturbation. For uncertainty, it uses Monte Carlo dropout with dropout probability $c=0.5$ and $T=20$ repeats, estimating predictive mean and variance from repeated stochastic forward passes. The paper reports that uncertainty changes are generally modest under most shifts, and only the strongest broadband noise causes a noticeable increase (Wagh et al., 2022).

The power-network studies validate realism operationally, statistically, and geometrically. Operational validation compares residential node voltages and edge power flows against an actual Blacksburg, Virginia, network; the 2022 paper reports that most synthetic residence voltages lie within $p_t(x)=p_s(\mathcal{T}(x)).$0 of the actual network’s voltages and that the edge-flow distributions have KL divergence $p_t(x)=p_s(\mathcal{T}(x)).$1. Statistical validation compares degree, hop, and reach distributions, with reported KL divergences $p_t(x)=p_s(\mathcal{T}(x)).$2, $p_t(x)=p_s(\mathcal{T}(x)).$3, and $p_t(x)=p_s(\mathcal{T}(x)).$4, respectively. Geometric validation uses a Hausdorff-distance-based comparison over rectangular grid cells (Meyur et al., 2022). The 2020 precursor likewise reports that synthetic voltages remain within $p_t(x)=p_s(\mathcal{T}(x)).$5 of actual voltages and that the simulated operating case stays within the ANSI C84.1 normal range of $p_t(x)=p_s(\mathcal{T}(x)).$6 to $p_t(x)=p_s(\mathcal{T}(x)).$7 pu (Meyur et al., 2020).

5. Assumptions, limits, and common misconceptions

A common misconception is that realism can be obtained by fitting a powerful joint model to the entire table or signal collection. The RCT results do not support that view. The paper states that CTGAN flexibility is not necessarily beneficial in an RCT context if temporal order is violated, and that simple regressions were sufficient and often superior to training more complex CTGANs at every post-treatment step (Petrakos et al., 29 Jan 2025).

A second misconception is that any synthetic perturbation is a satisfactory robustness benchmark. The EEG study argues against this explicitly: the setting is “realistic” not because it perfectly reproduces all deployment conditions, but because it targets a subset of plausible, domain-specific, clinically meaningful perturbations that are closer to real EEG variability than generic synthetic corruptions. The authors also note that they do not model all plausible sources of shift, especially physiological variability such as sleep state or other brain-state changes, and that the quantization model is a simplification (Wagh et al., 2022).

A third misconception is that a realistic synthetic network can be validated by abstract topology alone. The power-network papers reject this by embedding the network in real geography, using actual residences, roads, and substations, enforcing radiality and voltage limits, and assigning line types and impedances. At the same time, the current version uses only positive-sequence impedance and does not model the full range of three-phase transformer configurations and zero-sequence effects, so the networks are most suitable for balanced-load studies and symmetrical three-phase analysis rather than detailed transient or mixed-phase dynamic studies (Meyur et al., 2022).

The limitations are correspondingly domain-specific. The RCT framework is demonstrated on a single dataset, assumes a single-stage randomized trial with fixed treatment assignment, handles missingness mostly by omission in regression execution models, and generates synthetic samples of the same size as the original trial (Petrakos et al., 29 Jan 2025). The EEG framework does not assume access to external target-domain data, but that also means its target shifts are synthetic constructions rather than direct samples from deployment (Wagh et al., 2022). The power-network framework produces feasible but not necessarily optimal ensemble variants in length, and its realism remains tied to the quality of open geographic data and the adopted engineering abstractions (Meyur et al., 2022, Meyur et al., 2020).

6. Acronym collisions and adjacent literatures

The acronym “RDS” has established meanings that are unrelated to “realistic distribution setting.” In survey methodology, RDS usually denotes respondent-driven sampling. In that literature, realism has a different but related role: it enters as a critique of idealized recruitment assumptions. The 2012 paper “On the impossibility of constructing good population mean estimators in a realistic Respondent Driven Sampling model” argues that, under a realistic RDS sampling model with without-replacement sampling and differential recruitment, the inclusion probabilities required by the Horvitz–Thompson estimator cannot be determined from the sample information alone. The authors therefore conclude that, unless additional information about the underlying population network is obtained, it is hopeless to conceive of a general theory of population mean estimation from current RDS data (Guntuboyina et al., 2012).

Recent respondent-driven-sampling work continues this realism theme in a different direction. “Randomized Recruitment Driven Sampling” introduces researcher-controlled randomization into cellphone-based recruitment and reports less biased estimates and improved confidence interval coverage than traditional RDS in simulations and in a study of Bangladeshi garment workers during the COVID-19 pandemic (Visokay et al., 27 Feb 2026). “Inference from multivariate differential recruitment in respondent-driven sampling data” replaces uniform random recruitment with a covariate-weighted Markov chain that incorporates multiple simultaneous covariates, both categorical and continuous, and reports that MDR-adjusted estimators are most valuable when recruitment is both nonrandom and shaped by multiple covariates (Reinoso et al., 11 Apr 2026). These are not uses of “realistic distribution setting” as such, but they illustrate the same methodological tension between idealized inference and realistic data-generating mechanisms.

In stochastic systems, “RDS” also denotes random dynamical systems. For example, the analysis of Kalman filtering with intermittent observations studies random algebraic Riccati equations through the lens of random dynamical systems and Markov-Feller operators, establishing weak convergence to a unique invariant distribution under stochastic boundedness (0903.2890). This usage is terminologically unrelated and should not be conflated with realistic distribution settings.

Taken together, the literature indicates that a realistic distribution setting is best understood not as a single standardized framework, but as a design principle: preserve the actual structure of the process under study, whether that structure is temporal and causal, instrumentation-related, geographic and electrical, or otherwise domain-specific. The substantive value of the setting depends on whether the imposed realism is validated against the relevant distributional, operational, or deployment behavior of the real system (Petrakos et al., 29 Jan 2025, Wagh et al., 2022, Meyur et al., 2022, Meyur et al., 2020).