Summary Inference

• Summary inference comprises methods that compress data into essential summaries, preserving critical inferential content while reducing dimensionality.
• It is applied across fields like psychology, genomics, and astrophysics to facilitate population-level, likelihood-free, and decision-preserving analyses.
• Target-aware compression in summary inference optimizes trade-offs between computational cost and inferential accuracy using methods from inverse-variance weighting to neural learning.

Summary inference denotes a family of inferential procedures in which the primary inferential object is not the full raw dataset but a compressed representation of it: a subject-level effect estimate with its variance, a low-dimensional statistic, a summary-level association vector, a learned embedding, or a structure-preserving abstraction such as a summary graph. Across psychology, neuroscience, ABC and SBI, cosmology, genomics, network analysis, causal graph analysis, and NLP, the common aim is to preserve the information relevant to a specific inferential target while reducing dimensionality, computational cost, or interpretive burden (Dowding et al., 2017, Fearnhead et al., 2010, Zeng et al., 21 Apr 2025, Park et al., 18 Feb 2026).

1. Conceptual scope

In the literature, “summary inference” does not denote a single method. It instead refers to several closely related regimes in which inference is mediated by summaries rather than full observations. A useful taxonomy is the following.

Setting	Summary object	Inferential target
Hierarchical or nested data	Subject-level effect estimate plus within-subject variance	Population effect, standard errors, and hypothesis tests
Likelihood-free and simulation-based inference	Hand-crafted, learned, localized, or hybrid summary statistics	Posterior approximation or parameter estimation
Summary-data and structural inference	Marginal association statistics, external moments, embeddings, summary DAGs	Conditional dependence, causal effects, semantic fidelity, or decision faithfulness

A central distinction running through these literatures is between compression that is merely convenient and compression that is inferentially adequate. In nested-data analysis, the difference is between reducing each subject to one number and ignoring precision information, versus retaining both an estimate and its variance (Dowding et al., 2017). In ABC, the distinction appears as the difference between arbitrary summaries and posterior-mean-like summaries that are optimal under quadratic loss (Fearnhead et al., 2010). In recent hybrid methods, the goal is not simply to append neural features to existing summaries, but to train the learned component to capture only information complementary to the predefined statistic (Makinen et al., 2024, Makinen et al., 2024).

Taken together, these works suggest that summary inference is best understood as target-aware compression. The summary is judged not by brevity alone, but by whether it preserves the inferential content relevant to a specified task: population-level inference, posterior concentration, causal adjustment, or downstream decision support.

2. Nested and hierarchical data

A canonical form of summary inference arises with nested data, where many observations are collected within each subject and inference is required at the population level. In this setting, simply pooling raw observations across subjects is generally invalid because observations are not identically distributed across subjects, and it can either reduce power or create spurious significance. The baseline “naive summary-statistic approach” computes one effect estimate per subject and applies a group-level $t$-test to those $S$ summaries. This controls false positives under both fixed-effect and random-effects models, but it is suboptimal because it ignores within-subject precision and estimates group uncertainty from only the $S$ subject summaries (Dowding et al., 2017).

The sufficient-summary-statistic approach augments each subject summary $\hat\theta_s$ with an estimate of its within-subject variance and combines subjects through a weighted mean,

$$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$

with

$$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$

Under inverse-variance weighting,

$\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$

the estimator is minimum-variance unbiased provided the subject-level estimators are unbiased and the weights are independent of the subject effects. Equal-weight sufficient-summary inference can also improve on the naive $t$-test because it estimates uncertainty from within-subject variances rather than only from across-subject dispersion (Dowding et al., 2017).

The paper draws a sharp distinction between fixed-effect and random-effects formulations. Under fixed effects, all subjects share one common true effect and the null is strong: no effect in any subject. Under random effects, true subject effects vary with between-subject variance $\sigma_{\text{rand}}^2$, and the inferential target is the mean population effect. The recommended default is random-effects analysis, because ignoring $\sigma_{\text{rand}}^2$ when it exists yields underestimated variance and anti-conservative $S$0-values. The paper recommends the DerSimonian–Laird estimator for $S$1 when the number of subjects is moderate to large, while explicitly noting instability for small $S$2 (Dowding et al., 2017).

A further subtlety is that inverse-variance weighting can become biased when subject-level effect size and within-subject variance are associated. Simulation 3 in the paper shows that negative effect–variance association can inflate false positives and positive association can bias the estimate downward so severely that a positive true effect may yield a significant negative group effect. The recommended diagnostic is to check monotonic association, for example with Spearman rank correlation; if dependence is present, equal weighting or sample-size weighting should replace inverse-variance weighting (Dowding et al., 2017).

The framework extends beyond simple means. The paper gives step-by-step subject-level summaries and variance formulas for one-sample means, paired mean differences, Welch-type within-subject differences between unpaired conditions, AUC via Mann–Whitney $S$3, Fisher-transformed Pearson correlations, and OLS regression coefficients. In the EEG simulated-driving example with 18 drivers and about 200 clean emergency events per driver, strong positive correlations between subject-level mean differences and within-subject standard deviations made inverse-variance weighting likely biased, so equal-weight sufficient-summary inference was used instead. The resulting group $S$4-scores showed earlier emergency-related information in EEG than EMG, and equal-weight sufficient-summary inference was more powerful than the naive paired $S$5-test on subject means (Dowding et al., 2017).

3. Summary statistics in likelihood-free and simulation-based inference

In ABC and SBI, summary inference addresses a different bottleneck: the likelihood is unavailable or inconvenient, simulation is feasible, and inference must proceed through low-dimensional statistics $S$6. The foundational semi-automatic ABC result is that, for parameter estimation under quadratic loss, the optimal summaries are the posterior means $S$7. Since these are unavailable analytically in the likelihood-free setting, they are approximated by a regression stage on simulated $S$8 pairs, and the fitted predictors are then used as ABC summaries (Fearnhead et al., 2010). This converts summary construction from an ad hoc design problem into a supervised-learning problem tied to the inferential objective.

Subsequent work broadens this logic. Localization of projection-based summary reduction estimates the transformation $S$9 in a neighborhood of the observed data rather than globally over the prior-predictive space. The paper’s LocalProjectionOptimized algorithm chooses the neighborhood size and other hyperparameters by validation on pseudo-observed datasets, minimizing summed RMSE over validation posteriors. This is motivated by the observation that posterior inference only requires the projection to behave well locally near $S$0, and that global linear or PLS fits can fail when the relation between summaries and parameters is heterogeneous across parameter space (Sirén et al., 2019).

Neural summary learning pushes the same agenda further. The autoencoder approach of ENCA and INCA treats the encoder output $S$1 as the ABC summary, with the first $S$2 coordinates regularized as parameter regressors and the remaining $S$3 coordinates free to capture additional posterior-relevant structure. The crucial modification is to condition the decoder on the stochastic noise used to generate the training sample—explicitly in ENCA through $S$4, implicitly in INCA through replicate-specific weighting—which pressures the encoder to retain parameter-relevant information while discarding nuisance randomness. The paper explicitly argues that, for many scientific applications, one needs $S$5 summary statistics to represent posterior-relevant structure such as multimodality or order parameters (Albert et al., 2022).

A more recent development is hybrid summary construction. In the cosmological “Hybrid Summary Statistics” framework, a traditional summary $S$6 is kept fixed and a neural summary $S$7 is trained to extract only the information about parameters $S$8 that is missing beyond $S$9. The central information-theoretic identity is

$\hat\theta_s$0

so maximizing incremental information beyond $\hat\theta_s$1 is equivalent to maximizing the mutual information of the concatenated summary $\hat\theta_s$2. Two objective families are proposed: the Posterior Entropy objective

$\hat\theta_s$3

and a classifier-based Cross Entropy objective built from joint-versus-product-of-marginals discrimination. In sparse-simulation regimes, the learned hybrid summaries outperform handcrafted summaries alone, neural summaries alone, and post hoc concatenation, precisely because the neural component is optimized for complementarity rather than redundancy (Makinen et al., 2024).

The weak-lensing generalization of IMNNs formalizes the same idea through an information-update equation. If $\hat\theta_s$4 is a predefined summary with Fisher matrix $\hat\theta_s$5, and $\hat\theta_s$6 is a learned neural summary, then the hybrid summary $\hat\theta_s$7 has update

$\hat\theta_s$8

with $\hat\theta_s$9 and derivative correction $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$0 that explicitly subtracts parameter sensitivity already expressible through $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$1. This makes complementarity a property of the loss rather than an after-the-fact interpretation. In weak lensing, the learned summaries were optimized to complement angular power spectrum estimates and extracted at least $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$2 and up to $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$3 as much information as the angular power spectrum in all noise regimes reported (Makinen et al., 2024).

A different branch of the SBI literature treats the learned summary as an empirical-likelihood surrogate. In the Cressie–Read framework, observed and simulated samples are transformed into discrepancy-minimizing weights under moment restrictions, and the learned summary

$$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$4

is defined as a log ratio of those weight-based empirical models plus a quadratic penalty aligning moment parameters. This summary is then used in a Metropolis scheme in place of an exact log-likelihood. The same construction is extended to multiple simulation replications, conditional moment restrictions, empirically relevant subsets of the observed sample, and weakly dependent data through blockwise moments (Befekadu, 30 Jan 2026). This suggests a broader interpretation of summary inference in SBI: summaries can be low-dimensional statistics, but they can also be sample-dependent pseudo-likelihood objects.

4. Evaluating, validating, and stress-testing summaries

Once summaries are introduced, their adequacy becomes an inferential problem in its own right. A central criterion is mutual information. In cosmological summary evaluation, sufficiency is expressed as

$$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$5

and approximate sufficiency is operationalized by comparing candidate summaries through $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$6. Because the exact conditional $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$7 is unavailable, the paper uses the Barber–Agakov lower bound

$$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$8

with $$\hat{\theta}=\frac{\sum_{s=1}^S \alpha_s \hat{\theta}_s}{\sum_{s=1}^S \alpha_s},$$9 represented by a masked autoregressive flow. In a 21 cm reionization problem, this framework ranked the scattering transform above the power spectrum and bispectrum across all contamination regimes, while also showing that observational contamination monotonically reduces inferred information content (Sui et al., 2023).

A second evaluative question is whether different parts of a summary vector are mutually coherent under the assumed model. In regression-based likelihood-free inference, the summary vector $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$0 may contain conflicting information about a parameter $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$1. The paper adapts deletion–imputation tools from interpretable machine learning to estimate the subset posterior $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$2 without retraining the regression model, using

$$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$3

Conflict is then quantified by the maximum log relative belief

$$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$4

and calibrated under the conditional predictive distribution $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$5. The resulting diagnostics reveal, for example, conflicts between lower-tail and upper-tail summaries in stereological inference and between low-regime and high-regime summaries in the Ricker model (Mao et al., 2020).

Summary-based inference can also be performed directly from marginal dependence statistics. In summary-statistics knockoff inference, the available data are the marginal $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$6-scores and an estimated feature correlation matrix $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$7. GhostKnockoff is used to generate knockoff copies of summary statistics,

$$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$8

and a new multiple-knockoff FWER filter defines feature-level statistics $$\mathrm{Var}(\hat{\theta})=\frac{\sum_{s=1}^S \alpha_s^2 \mathrm{Var}(\hat{\theta}_s)}{\left(\sum_{s=1}^S \alpha_s\right)^2}.$$9 and $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$0 by comparing the original statistic to its knockoff copies. The stopping rule is calibrated through a negative binomial bound so that $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$1. The paper also gives an exact covariance decomposition reducing the cost of knockoff generation from $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$2 to $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$3 without sacrificing power or FWER control (Yu et al., 2023).

Taken together, these results suggest that summary inference is inseparable from summary validation. Compression does not merely raise the question “what is informative?” It also raises “what is redundant?”, “what conflicts under the model?”, and “what guarantees survive when only summary-level observables are available?”

5. Summary-data inference in biostatistics and physical science

In some literatures, the summaries are not handcrafted compressions of rich data but the only data available. Case-control logistic regression is a clear example. Internal retrospective sampling identifies the slope parameter $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$4 and the tilted intercept $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$5, but not the prospective intercept $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$6 or the prevalence $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$7. The paper restores identifiability by incorporating external summary information on marginal covariate moments $\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$8. The key identity is

$\alpha_s=\frac{1}{\mathrm{Var}(\hat\theta_s)},$9

where $t$0 and $t$1 are identifiable from cases and controls. An empirical-likelihood objective combines the internal case-control likelihood with a quadratic term for the external summary estimator $t$2, and the asymptotic covariance is minimized by choosing the weighting matrix $t$3, the covariance of $t$4 (Shi et al., 2024). Here summary inference is not a convenience; it is the mechanism by which otherwise unidentified parameters become estimable.

A related summary-data setting appears in nonlinear IV inference for TWAS. The model

$t$5

decouples a marginal causal effect $t$6 from an unknown nonlinear transformation $t$7, while still permitting the second stage to use only GWAS summary quantities

$t$8

The first-stage index $t$9 is estimated by sliced inverse regression from individual-level eQTL data, the causal component $\sigma_{\text{rand}}^2$0 is estimated by sparse summary-data IV regression robust to a sparse set of invalid instruments $\sigma_{\text{rand}}^2$1, and the nonlinear transformation is reconstructed by ratio-adjusted inverse regression. In the ADNI–IGAP application, the combined 2SIR procedure identified 18 putative AD genes, including APOE and TOMM40, and seven genes missed by linear 2SLS baselines (Dai et al., 2022).

In network analysis, the inferential target can itself be a network summary statistic of a latent graph. With a single observed network $\sigma_{\text{rand}}^2$2 corrupted by false negatives $\sigma_{\text{rand}}^2$3 and false positives $\sigma_{\text{rand}}^2$4, the paper first debiases the adjacency matrix by

$\sigma_{\text{rand}}^2$5

then spectrally truncates its eigendecomposition to form a denoised estimator $\sigma_{\text{rand}}^2$6. Summary statistics such as density, degree centrality, conductance, and eigenvector-centrality-related quantities are then estimated as $\sigma_{\text{rand}}^2$7 rather than $\sigma_{\text{rand}}^2$8. Because many such functionals are Lipschitz in $\sigma_{\text{rand}}^2$9, improved Frobenius risk for adjacency estimation translates directly into improved summary inference. In simulations and biological-network examples, density estimation improved by several orders of magnitude (Balachandran et al., 2013).

Gravitational-wave SBI supplies another example of likelihood-oriented summary data. For BNS signals, the raw frequency-domain data are too large for practical NPE, so the paper constructs summary statistics from relative binning. With reference waveform $\sigma_{\text{rand}}^2$0, bin-wise moments

$\sigma_{\text{rand}}^2$1

retain the information needed to approximate $\sigma_{\text{rand}}^2$2 and $\sigma_{\text{rand}}^2$3 under a polynomial model for the waveform ratio. This yields compression factors between 208 and 1236 relative to the raw frequency series. A neural posterior estimator trained on these summary data was validated against nested sampling on 1024 BNS injections, with well-calibrated posteriors and Jensen–Shannon divergences consistent with numerical noise for most parameters (Iwaya et al., 24 Apr 2026). A plausible implication is that, in physical SBI, the most effective summaries are often those designed around the algebra of the likelihood rather than generic feature extraction.

6. Decision-preserving and structure-preserving summarization

A different but related line of work treats summaries as objects that must preserve downstream decisions or structural semantics. In decision-focused summarization, the goal is not to maximize textual salience but to choose an extractive summary $\sigma_{\text{rand}}^2$4 such that a predictive model $\sigma_{\text{rand}}^2$5 makes approximately the same decision from the summary as from the full text. DecSum formalizes this through decision faithfulness,

$\sigma_{\text{rand}}^2$6

decision representativeness, defined via Wasserstein distance between sentence-level prediction distributions, and a textual non-redundancy term based on sentence-embedding cosine similarity. On a Yelp benchmark where the first ten reviews were summarized to support prediction of the average rating after fifty reviews, DecSum substantially outperformed text-only and explanation-style baselines in decision faithfulness and representativeness, and was the only method that enabled human participants to perform above random chance in the pairwise future-rating task (Hsu et al., 2021). This makes explicit a strong notion of inferential adequacy: the summary is good if the downstream judgment remains recoverable.

LLM summary evaluation introduces yet another form of summary inference. LIDS represents the source text $\sigma_{\text{rand}}^2$7 and repeated LLM-generated summaries $\sigma_{\text{rand}}^2$8 by BERT embedding matrices and SVD-based latent layers, then defines a layered direction vector

$\sigma_{\text{rand}}^2$9

and scores each summary by the maximum absolute cosine similarity

$S$00

Repeated prompting provides an empirical distribution of summary quality rather than a single deterministic score, and SOFARI plus Benjamini–Hochberg is used to identify layer-specific keywords with FDR control. On the Utah article, GPT-5 summaries had LIDS scores clearly above naive and random-topic baselines, and LIDS correlated strongly with human judgments while remaining substantially cheaper than BERTScore (Park et al., 18 Feb 2026). Here summary inference concerns semantic preservation, uncertainty across repeated summaries, and interpretable latent themes.

Structure-preserving summarization is even more explicit in causal DAG summarization. The summary DAG $S$01 is obtained by node contraction, but its semantics are defined through a canonical causal DAG over the original variables. The paper introduces $S$02-separation as the appropriate conditional-independence notion for summary DAGs and proves soundness and completeness statements linking $S$03-separation to $S$04-separation in the canonical DAG. The central inferential claim is that summary DAGs can be used directly for causal inference: node contraction behaves like conservative edge addition, so identifiability may be weakened but should not become spuriously stronger. The CaGreS algorithm greedily minimizes the number of added edges in the canonical DAG, and empirical studies on six datasets show that the resulting summary DAGs preserve causal-effect intervals and are more robust to misspecification than general-purpose graph summarizers (Zeng et al., 21 Apr 2025).

Taken together, these works suggest that summary inference is no longer confined to classical statistics. It now includes decision-preserving compression, uncertainty quantification for generated summaries, and graph abstractions that remain usable for do-calculus. Across all of these settings, the recurring question is the same: what minimal representation preserves the inferential content that actually matters?