Entropic Projection Alignment (EPA)

• Entropic Projection Alignment (EPA) is a framework that uses entropic regularization to importance weight source data by carefully matching selected moments with those of the target domain.
• It employs a closed-form solution derived from Lagrangian optimization that minimizes KL divergence, ensuring low variance and robustness in model performance estimation.
• EPA simultaneously estimates target risk, explains distribution shifts by identifying key features, and adapts models through efficient, entropy-regularized corrections.

Entropic Projection Alignment (EPA) is, in the formulation of "Entropic Projection Alignment: Estimating, Explaining, and Improving Model Performance Under Distribution Shift" (Amoukou et al., 29 May 2026), a unified framework for importance-weighting the source distribution so that it both matches key statistics of an unlabeled target domain and stays as close as possible to the original source. It is designed to address three tasks under distribution shift: estimating a model’s performance on an unlabeled target domain, explaining the shift by identifying the features responsible, and improving target-domain performance. The method aligns the source distribution to the target by matching carefully selected moments while simultaneously minimising the KL divergence from the source, which yields a unique closed-form solution for importance weights and robustness through implicit variance control (Amoukou et al., 29 May 2026).

1. Distribution-shift setting and objectives

EPA is posed on a source joint law $P$ on $\mathcal{X}\times\mathcal{Y}$ and a target joint law $Q$. Under covariate or joint shift, the stated assumption is that $Q_{Y|X}=P_{Y|X}$ but $P_X\neq Q_X$, or more generally that there is some hypothesis in $H$ that does well under both $P$ and $Q$ (Amoukou et al., 29 May 2026). The observed data consist of a labeled source sample $\{(X_i,Y_i)\}_{i=1}^n\sim P$ and only unlabeled target samples $\{X_j^{(t)}\}_{j=1}^m\sim Q$.

The central objective is to construct nonnegative weights $\mathcal{X}\times\mathcal{Y}$0 on source examples so that the weighted source distribution $\mathcal{X}\times\mathcal{Y}$1 matches key target statistics, $\mathcal{X}\times\mathcal{Y}$2 remains close to $\mathcal{X}\times\mathcal{Y}$3, and downstream quantities such as the target risk

$\mathcal{X}\times\mathcal{Y}$4

can be estimated or optimized via

$\mathcal{X}\times\mathcal{Y}$5

In the description of (Amoukou et al., 29 May 2026), this produces three outputs: unbiased, low-variance estimates of a model’s target risk; an explainer for which features drive the shift; and a recipe for model correction under shift.

A defining feature of EPA is that it does not attempt unrestricted source-to-target matching. Instead, it matches only carefully selected moments through a feature map $\mathcal{X}\times\mathcal{Y}$6, for example quantile-bins of important scalar scores or features (Amoukou et al., 29 May 2026). This is the mechanism by which the framework couples estimation, explanation, and adaptation.

2. Entropic projection formulation and closed-form weights

EPA is formulated as an $\mathcal{X}\times\mathcal{Y}$7-projection, or entropic projection, of the source onto the set of distributions that match target moments (Amoukou et al., 29 May 2026). With $\mathcal{X}\times\mathcal{Y}$8, the optimization problem is

$\mathcal{X}\times\mathcal{Y}$9

In density-ratio form, with $Q$0, this becomes

$Q$1

subject to

$Q$2

The Lagrangian is given as

$Q$3

Stationarity in $Q$4 yields

$Q$5

hence

$Q$6

After enforcing normalization, one obtains the stated closed form

$Q$7

The dual variables are determined by the convex dual problem

$Q$8

whose gradient vanishes exactly when

$Q$9

Standard convex solvers, specifically L-BFGS, are used to find $Q_{Y|X}=P_{Y|X}$0 (Amoukou et al., 29 May 2026).

This formulation is the technical core of EPA. The importance weights are exponential-family weights induced by the moment constraints, and the optimization is convex in the dual variables. A plausible implication is that EPA’s practical behavior is driven less by unrestricted density-ratio estimation than by the selection of the moment map $Q_{Y|X}=P_{Y|X}$1.

3. Variance control and domain-adaptation guarantees

EPA’s robustness is tied to the KL objective. Since

$Q_{Y|X}=P_{Y|X}$2

the solution maximizes the entropy of $Q_{Y|X}=P_{Y|X}$3 subject to the constraints (Amoukou et al., 29 May 2026). In the paper’s terminology, this encourages weights near uniform, preventing extreme weights and controlling the variance of

$Q_{Y|X}=P_{Y|X}$4

The stated conditional variance bound is Theorem 3.1: for any bounded loss $Q_{Y|X}=P_{Y|X}$5, conditionally on the features,

$Q_{Y|X}=P_{Y|X}$6

and

$Q_{Y|X}=P_{Y|X}$7

where $Q_{Y|X}=P_{Y|X}$8. EPA minimizes $Q_{Y|X}=P_{Y|X}$9, thereby minimizing this upper bound (Amoukou et al., 29 May 2026).

The paper also places EPA within domain adaptation theory. Defining the $P_X\neq Q_X$0-discrepancy between $P_X\neq Q_X$1 and $P_X\neq Q_X$2 as

$P_X\neq Q_X$3

the standard Ben-David et al. bound gives, for any $P_X\neq Q_X$4,

$P_X\neq Q_X$5

where $P_X\neq Q_X$6 is the combined risk of the best cross-domain hypothesis (Amoukou et al., 29 May 2026).

EPA’s central theoretical claim is that moment matching can reduce $P_X\neq Q_X$7 to zero when $P_X\neq Q_X$8 aligns with the disagreement space:

$P_X\neq Q_X$9

The examples given are specific. For threshold classifiers in $H$0, binning of $H$1 into quantiles approximates any indicator disagreement. For linear regression with squared loss, if $H$2 includes degree-2 monomials then any difference of two linear predictors, described as a quadratic function, lies in $H$3. For trees, one could include one indicator per tree-leaf region to span all disagreements, though in practice the paper approximates this with low-cardinality quantile bins of tree scores (Amoukou et al., 29 May 2026).

The significance of these results is narrow but precise: EPA’s guarantees are framed around matching the moments necessary for low discrepancy, rather than full density-ratio recovery. This suggests a selective notion of alignment in which correctness depends on the relation between $H$4 and the hypothesis class $H$5.

4. Operational procedures: estimation, explanation, and adaptation

The implementation described in (Amoukou et al., 29 May 2026) is organized into four algorithms.

Algorithm 1, Core EPA, takes source data, target data, and a feature map $H$6. It computes

$H$7

solves

$H$8

via L-BFGS, and sets

$H$9

The stated cost per iteration is $P$0 to evaluate $P$1 and $P$2, so the total is $P$3 with $P$4–$P$5, where $P$6.

Algorithm 2 estimates target risk. It uses the same $P$7 augmented with model scores $P$8 binned into $P$9 bins, computes EPA weights, and returns

$Q$0

Algorithm 3 explains the shift. For each candidate feature-subset $Q$1, it defines $Q$2 to include only those feature-bins, runs Core EPA, computes histogram-KL between reweighted source and target marginals, and picks the minimal-cardinality $Q$3 achieving near-zero KL (Amoukou et al., 29 May 2026). This makes the explanation problem a constrained search over feature subsets under the same entropic-projection mechanism.

Algorithm 4 adapts the model. It runs EPA with $Q$4 combining features and current model predictions, obtains weights $Q$5, and then trains a lightweight correction via weighted boosting, for example XGBoost with instance weights $Q$6, requiring only $Q$7 per boosting round (Amoukou et al., 29 May 2026).

Across these modes, the same weighting mechanism is reused with different feature maps. That reuse is one of the paper’s unifying design choices: risk estimation, explanation, and adaptation are not separate estimators, but separate instantiations of the same entropic-projection program.

5. Empirical evaluation

The empirical study in (Amoukou et al., 29 May 2026) covers three shift scenarios:

• Sparse Covariate Shift: Adult, HELOC, and NHANES I, generating 292 one-feature splits.
• Sparse Joint Shift: including label undersampling and oversampling.
• Natural Shift: Folktables ACSIncome across states and years.

The baselines are SEES-C and SEES-D from Chen et al. 2022, Kernel Mean Matching (KMM), domain-classifier reweighting (covcolour), uniform calibration (caluni), and target calibration (produni) (Amoukou et al., 29 May 2026).

The reported metrics are estimation inaccuracy $Q$8, detection power for harmful shifts and false-positive rate, explanation accuracy or feature correlation in sparse-shift cases, model-improvement measured as percentage reduction in target error and proportion of shifts improved (I-prop), and runtime per weighting (Amoukou et al., 29 May 2026).

The key findings are numerically specific. EPA’s estimation error was 30–70% lower than all baselines, especially in harmful bins B3 and B4. Detection power reached $Q$9–$\{(X_i,Y_i)\}_{i=1}^n\sim P$0 versus $\{(X_i,Y_i)\}_{i=1}^n\sim P$1–$\{(X_i,Y_i)\}_{i=1}^n\sim P$2 for competitors. In explanation, EPA recovered the true splitting feature with up to 60% higher accuracy. In adaptation, EPA-weighted boosting reduced target error by 6–25% on harmful shifts, matching or outperforming the ideal produni, with $\{(X_i,Y_i)\}_{i=1}^n\sim P$3 versus $\{(X_i,Y_i)\}_{i=1}^n\sim P$4–$\{(X_i,Y_i)\}_{i=1}^n\sim P$5 for others. Runtime was approximately $\{(X_i,Y_i)\}_{i=1}^n\sim P$6 seconds per shift for EPA, compared with $\{(X_i,Y_i)\}_{i=1}^n\sim P$7 seconds for SEES-D, $\{(X_i,Y_i)\}_{i=1}^n\sim P$8 seconds for SEES-C, and approximately $\{(X_i,Y_i)\}_{i=1}^n\sim P$9 seconds for KMM (Amoukou et al., 29 May 2026).

Within the scope of those experiments, the paper’s summary claim is that EPA offers a principled, closed-form, entropy-regularized importance weighting that matches only those moments necessary to guarantee low discrepancy while simultaneously controlling variance. The empirical profile is therefore not limited to predictive accuracy; it also includes explanatory sparsity and computational efficiency.

6. Related entropic formulations and nomenclature overlap

The phrase “Entropic Projection Alignment” also appears in distinct but related entropic-regularized settings, and this creates a genuine terminological overlap.

In "Entropic Optimal Transport Eigenmaps for Nonlinear Alignment and Joint Embedding of High-Dimensional Datasets" (Landa et al., 2024), EPA is presented as a method for simultaneously aligning and embedding two high-dimensional datasets into a common low-dimensional space. The core object is the entropic-regularized optimal-transport plan between two datasets, with ground-cost matrix

$\{X_j^{(t)}\}_{j=1}^m\sim Q$0

and solution

$\{X_j^{(t)}\}_{j=1}^m\sim Q$1

computed by Sinkhorn iterations. The method then uses the nontrivial singular vectors of $\{X_j^{(t)}\}_{j=1}^m\sim Q$2 to embed the datasets jointly, yielding what the paper describes as an inter-data Laplacian eigenmap when $\{X_j^{(t)}\}_{j=1}^m\sim Q$3 and an inter-data diffusion map when $\{X_j^{(t)}\}_{j=1}^m\sim Q$4 (Landa et al., 2024). Under a latent-manifold model with arbitrary shifts, scalings, nuisances, and noise, the paper states that EPA recovers the shared manifold geometry via convergence of the transport plan to a population kernel and of its spectrum to that of the Laplace–Beltrami operator.

A further related use appears in "Entropic optimal transport is maximum-likelihood deconvolution" (Rigollet et al., 2018). There, entropic optimal transport is shown to be algebraically equivalent to maximum-likelihood estimation in Gaussian deconvolution, and EPA is defined as projection of an empirical target measure $\{X_j^{(t)}\}_{j=1}^m\sim Q$5 onto a model class $\{X_j^{(t)}\}_{j=1}^m\sim Q$6 by minimizing the entropic OT cost

$\{X_j^{(t)}\}_{j=1}^m\sim Q$7

The corresponding optimization is

$\{X_j^{(t)}\}_{j=1}^m\sim Q$8

typically with $\{X_j^{(t)}\}_{j=1}^m\sim Q$9 and $\mathcal{X}\times\mathcal{Y}$00 in the deconvolution interpretation (Rigollet et al., 2018). Sinkhorn iterations provide the inner entropic-OT computation.

These usages are mathematically adjacent but operationally distinct. The distribution-shift EPA of (Amoukou et al., 29 May 2026) is a KL-minimizing importance-weighting method with moment constraints on source samples; the alignment-and-embedding EPA of (Landa et al., 2024) is an entropic-OT spectral construction on paired datasets; and the EPA of (Rigollet et al., 2018) is a model-class projection under entropic OT with a maximum-likelihood deconvolution interpretation. The common thread is the use of entropic regularization and projection-like alignment, but the optimized objects, guarantees, and applications differ. This suggests that “EPA” is best interpreted as a family resemblance across entropic projection and alignment methods rather than a single universally fixed formalism.