Augmented Statistics: Enhanced Inference & Data Fusion

• Augmented statistics are an advanced approach that integrates auxiliary information, such as latent variables and invariant features, into conventional statistical measures.
• This technique enhances inferential procedures by incorporating methods like latent-data conditioning, learned residual corrections, and prediction-based sampling.
• Empirical studies across federated learning, cosmology, and simulation-based inference demonstrate improved estimation precision, robustness, and assumption validation.

Searching arXiv for recent and directly relevant papers on “augmented statistics” and adjacent usages of the term. Search query: augmented statistics arXiv summary statistics bootstrap data augmentation federated learning Searching arXiv… Augmented statistics denotes a heterogeneous family of constructions in which a statistic, summary, or inferential object is enriched with auxiliary information rather than treated as a fixed reduction of the data. In recent work, augmentation has meant conditioning on robust but insufficient summaries through latent data, mixing per-image channel moments across clients, attaching learned residual summaries to handcrafted summaries, extending confidence procedures by approximately invariant transformations, adding latent coordinates to correlation functions, and correcting classical estimators with predictions or selective labeling rules (Luciano et al., 2023, Lewy et al., 2022, Makinen et al., 2024, Huang, 8 Jun 2026, Bianchi, 28 May 2026, Wang et al., 12 May 2026).

1. Semantic scope and unifying pattern

The term is used in several technically distinct senses. In some papers, the augmented object is the statistic itself: StatMix defines an augmented statistic as the pair of first- and second-order pixel-value statistics, namely per-channel mean and standard deviation, and then transfers those six numbers between images. In other papers, the statistic is lifted by latent variables or auxiliary coordinates: the augmented correlation function adds a latent variable $\lambda$ to the two-point function, while Insufficient Gibbs Sampling augments the observed summary $S=T(X)$ with a latent sample $X$. A third usage augments the inferential procedure around a statistic: DAB augments a statistic $f(X)$ by transformed replicates $f(\Phi_b(X))$, and active inference for $U$-statistics augments inverse-probability weighting with machine-learning predictions. Hybrid Summary Statistics and SSAR further extend the idea to learned residual summaries and statistical-space priors (Lewy et al., 2022, Luciano et al., 2023, Bianchi, 28 May 2026, Huang, 8 Jun 2026, Wang et al., 12 May 2026, Makinen et al., 2024, Choi et al., 2024).

Formulation	Augmented object	Stated objective
"Insufficient Gibbs Sampling" (Luciano et al., 2023)	latent data $X$ given $S=T(X)$	exact Bayesian inference from robust but insufficient statistics
"StatMix" (Lewy et al., 2022)	$S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$	improve FL using image statistics
"Hybrid Summary Statistics" (Makinen et al., 2024)	$z(d)=[t(d),s(d)]$	maximize information extraction under sparse simulations
"Data augmented bootstrap" (Huang, 8 Jun 2026)	list $S=T(X)$0	confidence intervals from approximate invariance
"Learning U-Statistics with Active Inference" (Wang et al., 12 May 2026)	augmented IPW $S=T(X)$1-statistic	improve estimation efficiency under a labeling budget
"Augmented Correlation Functions for Spectroscopic Galaxy Surveys" (Bianchi, 28 May 2026)	$S=T(X)$2	isolate clustering properties averaged out by standard two-point statistics

This suggests a common structural motif: augmentation does not replace the original statistic, but embeds it in a richer object whose extra coordinates encode latent data, style information, invariance, learned residual information, or sampling design.

2. Latent-data conditioning and auxiliary-variable inference

A canonical latent-data formulation appears in "Insufficient Gibbs Sampling" (Luciano et al., 2023). There the observed object is a robust but insufficient summary $S=T(X)$3, with examples including $S=T(X)$4, $S=T(X)$5, or a collection of quantiles. The joint posterior is written

$S=T(X)$6

and inference proceeds by a Gibbs sampler on $S=T(X)$7-space. The $S=T(X)$8-step samples from $S=T(X)$9 by exact truncation, Metropolis-Hastings on order statistics, or pairwise updates preserving the median+MAD configuration; the $X$0-step samples from $X$1, often in conjugate form. A by-product is Bayes-factor estimation from observed statistics via bridge sampling. The paper reports recovery of the exact posterior in a Gaussian example with observed $X$2, a tighter posterior than ABC for Cauchy location-scale inference under the same CPU time, consistency in Normal-versus-Laplace model choice when $X$3 is observed, and a real-data application to 2020 French commune incomes in which insufficient bridge sampling yields $X$4-BF $X$5 in favor of Gamma.

A broader auxiliary-variable perspective is given by AXDA, which introduces splitting variables $X$6 and a smoothing kernel $X$7 to define

$X$8

The approximation is asymptotically exact as $X$9, with non-asymptotic bounds such as $f(X)$0 and total-variation control under Lipschitz or convexity assumptions. The framework supports Gibbs sampling, mean-field variational Bayes, quadratic-penalty or ADMM-style optimization, and EM-type algorithms, and is presented as a systematic alternative to bespoke exact augmentations (Vono et al., 2019).

An older but related line is parameter expansion in testing. There the original sufficient statistic $f(X)$1 is extended to $f(X)$2 under an expanded family $f(X)$3, and the activated component $f(X)$4 strictly reduces the upper bound on the sum of type I and type II error probabilities when the conditional densities in $f(X)$5 differ under the competing hypotheses. The gain is summarized by a scalar measure

$f(X)$6

which quantifies the effect of activating the additional component in the sufficient statistic (Yatracos, 2015).

3. Statistic-space augmentation in machine learning

In federated learning, "StatMix" uses statistics themselves as the augmentation medium. For each image $f(X)$7, the extracted statistic is

$f(X)$8

A local image is normalized by its own channel statistics and then re-stylized with a source statistic set from the global pool:

$f(X)$9

Only a Bernoulli switch with probability $f(\Phi_b(X))$0 decides whether a batch is augmented. The method leaves FedAvg unchanged and exchanges only 6 real numbers per image. In the main experiments $f(\Phi_b(X))$1. On CIFAR-10 with 5 nodes and DLA, accuracy moves from $f(\Phi_b(X))$2 to $f(\Phi_b(X))$3; with 10 nodes it moves from $f(\Phi_b(X))$4 to $f(\Phi_b(X))$5. On CIFAR-100 with PreActResNet-18 and 10 nodes, accuracy moves from $f(\Phi_b(X))$6 to $f(\Phi_b(X))$7, whereas with 50 nodes it decreases from $f(\Phi_b(X))$8 to $f(\Phi_b(X))$9 (Lewy et al., 2022).

"Hybrid Summary Statistics" augments a fixed, hand-crafted summary $U$0 with a learned neural residual summary $U$1 and forms

$U$2

The design goal is to learn only the information not already captured by $U$3 by maximizing mutual information with the parameter:

$U$4

Two loss formalisms are introduced: the posterior-entropy loss and the cross-entropy classification loss. In the 21 cm application, $U$5 and $U$6; in weak lensing, $U$7 and $U$8. With 10,000 21 cm simulations, the power-spectrum baseline has $U$9, while the hybrid EPE variant yields $X$0 and posterior area ratio $X$1. In weak lensing with 500 simulations, the $X$2 baseline has $X$3 for $X$4, CNN-only has $X$5, and Hybrid (EPE) has $X$6 with area ratio $X$7 (Makinen et al., 2024).

SSAR, or Statistical-space Augmented Representation, constructs a time-indexed graph $X$8 from sliding-window statistics. For each directed pair $X$9,

$S=T(X)$0

where $S=T(X)$1 is chosen from Pearson, Spearman, Kendall, Granger causality, mutual information, or transfer entropy. The augmented representation is $S=T(X)$2, and no extra trainable embedding layers are required for the statistics. On Data Set 1, the best SSAR result is MSE $S=T(X)$3 (TE, $S=T(X)$4) versus best baseline $S=T(X)$5 (LSTM). On Data Set 2, median MSE for SSAR variants is $S=T(X)$6 versus $S=T(X)$7 for GRU and $S=T(X)$8 for DLinear. Paired $S=T(X)$9-tests give $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$0 in all comparisons, and an $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$1-test reports $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$2 with $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$3 (Choi et al., 2024).

These examples use augmentation in statistic-space rather than direct perturbation of raw observations. The resulting object is often lower-dimensional, modular, and explicitly structured around invariance, style, regime, or residual information.

4. Transformation-, prediction-, and sampling-augmented inference

DAB recasts confidence-interval construction in terms of approximately invariant transformations. With a statistic $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$4 and transformations $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$5, the augmented list is

$S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$6

Exact exchangeability of this list would yield a uniform rank pivot; DAB relaxes this to approximate exchangeability and approximate invariance, measured in Kolmogorov distance. The conditional approximate invariance quantity is

$S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$7

The associated $S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$8-value is

$S(x)=\{\mu_c(x),\sigma_c(x)\}_{c=1}^3$9

and confidence sets are obtained by inversion. DAB recovers classical bootstrap, split conformal prediction, wild bootstrap for MMD-$z(d)=[t(d),s(d)]$0, and SymmPI as special cases, and its coverage theorems interpolate between finite-sample and asymptotic validity. Under Gaussian universality, conditional mean and variance matching can make $z(d)=[t(d),s(d)]$1 (Huang, 8 Jun 2026).

Prediction-augmented inference appears in PART and PAQ. Given a small labeled sample $z(d)=[t(d),s(d)]$2, a much larger unlabeled sample $z(d)=[t(d),s(d)]$3, and a pretrained predictor $z(d)=[t(d),s(d)]$4, PART replaces the global residual correction of PPI by a tree-based mixture of local corrections. With residuals $z(d)=[t(d),s(d)]$5, candidate splits are chosen by minimizing

$z(d)=[t(d),s(d)]$6

and the mean estimator is

$z(d)=[t(d),s(d)]$7

The paper derives asymptotic normality, a consistent variance estimator, and Wald-type confidence intervals. In the infinite-depth limit, PAQ yields a nearest-neighbor quadrature estimator with variance $z(d)=[t(d),s(d)]$8, improving on the $z(d)=[t(d),s(d)]$9 rate of PPI and PPI++, and empirically yields often a $S=T(X)$00-$S=T(X)$01 width reduction compared to PPI++ (Kher et al., 19 Oct 2025).

Active inference for $S=T(X)$02-statistics augments inverse-probability weighting by machine-learning predictions. For a symmetric kernel $S=T(X)$03 of order $S=T(X)$04, a predictor $S=T(X)$05, and Bernoulli labeling indicators $S=T(X)$06, the augmented IPW estimator combines a full-sample plug-in term with a Horvitz-Thompson correction on sampled tuples. The optimal sampling design is characterized by

$S=T(X)$07

where $S=T(X)$08 is a first-order residual score. The framework gives unbiasedness, asymptotic normality, and valid confidence intervals, and extends to $S=T(X)$09-statistic-based empirical risk minimization. Reported gains include $S=T(X)$10 larger effective sample size than the classical uniform-IPW baseline for Gini-index estimation, about $S=T(X)$11 fewer labels for a Wilcoxon signed-rank target, and $S=T(X)$12-$S=T(X)$13 savings for a third-order $S=T(X)$14-statistic in simulation (Wang et al., 12 May 2026).

5. Scientific realizations and empirical behavior

In small-sample tabular prediction, synthetic augmentation is presented as a form of augmented statistics through bootstrap resampling, sequential decision trees, Bayesian networks, CTGAN, and TVAE. A decision-support logistic model uses base sample size $S=T(X)$15, imbalance factor IF, degrees of freedom DF, and baseline AUC $S=T(X)$16 to predict whether augmentation is useful. On seven real health datasets, augmentation increases AUC by between $S=T(X)$17 and $S=T(X)$18, with average relative improvement $S=T(X)$19 and one-tailed $S=T(X)$20 versus baseline. Augmentation AUC is higher than resampling-only AUC with $S=T(X)$21, and diversity of augmented data is higher than diversity of resampled data with $S=T(X)$22 (Liu et al., 30 Jan 2025).

In cosmology, the augmented correlation function extends the standard two-point statistic by a latent coordinate:

$S=T(X)$23

In redshift space, the practical estimator is the Landy-Szalay extension

$S=T(X)$24

The proof-of-concept latent variable is the pairwise gradient of $S=T(X)$25, producing $S=T(X)$26 for infalling and $S=T(X)$27 for outflowing pairs. Quantile decomposition separates these regimes and yields Fisher-forecast improvements that are largest for $S=T(X)$28 and $S=T(X)$29, at roughly $S=T(X)$30-$S=T(X)$31, with more modest improvements on $S=T(X)$32, $S=T(X)$33, and $S=T(X)$34, and a $S=T(X)$35 gain on $S=T(X)$36 (Bianchi, 28 May 2026).

In small-sample linear regression, GP-MEVT combines a Gaussian Process with a Modified Extreme Value Theorem to generate augmented observations beyond the observed predictor range while preserving linear structure and controlled variability. The method is tested across $S=T(X)$37 and $S=T(X)$38. For $S=T(X)$39, assumption-satisfaction rates are $S=T(X)$40 for GP-MEVT, $S=T(X)$41 for bootstrap, and $S=T(X)$42 for bootstrap+noise. In a real-world dataset subsampled to $S=T(X)$43, the overall pass rate is $S=T(X)$44 for GP-MEVT versus $S=T(X)$45 and $S=T(X)$46 for the bootstrap alternatives; the abstract summarizes this as a $S=T(X)$47 assumption satisfaction rate compared to $S=T(X)$48 and $S=T(X)$49 (Salay et al., 16 Jun 2026).

Across these applications, the augmented object differs sharply by domain: latent robust summaries in Bayesian inference, image moment transfer in federated learning, neural residual summaries in simulation-based inference, graph-valued statistical priors in time series, latent pair labels in galaxy clustering, and synthetic samples or corrected estimators in small-sample statistics. The empirical commonality is not a single metric but a repeated claim that carefully chosen augmentation can improve inferential precision, robustness, or assumption satisfaction.

6. Limitations, misconceptions, and open problems

A central misconception is that augmented statistics are synonymous with sufficient statistics. "Insufficient Gibbs Sampling" is built precisely for the opposite case: robust summaries such as median, MAD, and IQR are not sufficient, and their inefficiency implies a posterior less concentrated than the full-data posterior. The same work emphasizes identifiability issues when the statistic dimension is too small, high-variance bridge sampling under wildly varying likelihoods, slow constrained mixing in low-probability regions, and the fact that when $S=T(X)$50 is not discriminative across models, $S=T(X)$51 need not converge to the “true” model (Luciano et al., 2023).

A second misconception is that augmentation is simply data duplication. In StatMix, the exchange unit is not the image but the six-number statistic $S=T(X)$52; in Hybrid Summary Statistics, augmentation means adding a learned low-dimensional residual summary to a fixed hand-crafted summary; in DAB, it means augmenting a rank pivot with approximately invariant transformations; in PART and active $S=T(X)$53-statistics, it means correction of classical estimators by predictions and sampling design rather than creation of synthetic records. This suggests that the operative question is not whether data are increased, but which inferential object is enriched and under what structural assumption.

A third misconception is that more augmentation is always better. StatMix reports that on CIFAR-10 any $S=T(X)$54 yields a $S=T(X)$55-$S=T(X)$56 pp gain, but very large $S=T(X)$57 or $S=T(X)$58 collapses accuracy to $S=T(X)$59; for CIFAR-100, smaller $S=T(X)$60 is optimal. In small tabular datasets, no specific generative model consistently outperformed the others, and augmentation can be negligible or negative when $S=T(X)$61, baseline AUC is high, or degrees of freedom are low. DAB notes that weak approximate invariance can make intervals conservative, while the augmented-correlation paper explicitly treats its Fisher improvements as indicative because of the exploratory nature of the analysis and the limitations of Fisher forecasts and simulations (Lewy et al., 2022, Liu et al., 30 Jan 2025, Huang, 8 Jun 2026, Bianchi, 28 May 2026).

Privacy and validity also remain qualified rather than automatic. StatMix lowers communication to 6 floats per image and argues that privacy risk is dramatically lower than sending raw or mixed pixels, yet it does not provide a formal DP guarantee. DAB provides validity under explicit approximate-invariance conditions, but not an efficiency oracle for choosing augmentations. Active $S=T(X)$62-statistics preserve valid inference under sampling-rule regularity, and PART/PAQ derive asymptotic guarantees under residual and smoothness assumptions; those assumptions are part of the method, not incidental details.

The literature therefore does not present augmented statistics as a single doctrine. It presents a family of techniques for retaining or recovering information that a conventional statistic, sampling design, or estimator would discard. The main open question, implicit across the surveyed works, is selection: which augmentation target, auxiliary variable, latent coordinate, prediction correction, or invariance class best matches a given scientific problem.