acceptodds
Under review as a conference paper at ICLR 2027

Semiparametric Efficiency for Entropic Optimal Transport under Structured Distribution Shift

Abstract

Entropic optimal transport (EOT) is widely used to compare distributions and learned representations, but its empirical estimate can be noisy when target data are scarce. We study whether a low-dimensional structural relation between source and target distributions can improve estimation of a fixed-regularization EOT functional. Under an exponential density-ratio model with a nonparametric baseline, we derive the canonical gradient and an exact efficiency identity: the attainable variance reduction is characterized by a weighted projection residual of the unrestricted transport influence contrast. We construct an empirical-likelihood estimator, DRM-EOT, that attains the semiparametric efficiency bound on compact support, together with consistent analytic variance estimation and a valid stratified refit bootstrap. The same residual also governs first-order sensitivity to local misspecification, yielding an explicit efficiency–robustness trade-off. Controlled simulations and exact semi-synthetic shifts on OfficeHome representations confirm the predicted variance gains. Under natural OfficeHome shifts, low-dimensional DRM can reduce target-scarce RMSE mainly through bias reduction, while increasing model complexity exposes separation and persistent misspecification bias. These results distinguish exact-model efficiency from structural regularization and characterize when information sharing across domains improves EOT estimation.

Primary area:
learning theory
Keywords:
entropic optimal transport, semiparametric efficiency, density ratio models, distribution shift, representation shift, empirical likelihood, statistical inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_k7XxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the estimation of entropic optimal transport between two distributions. It asks whether using the relationship between the distributions can improve the accuracy of EOT estimation compared to estimating the two distributions separately. The authors consider this question for distributions linked by a density ratio model. The DRM-EOT method proposed by the authors estimates this density ratio and uses the pooled sample with reweighting to estimate both distributions. EOT is then computed between them. The authors derive an exact formula for the reduction in asymptotic variance compared to the standard empirical estimator and prove that, under the assumptions, their method reaches the theoretical efficiency bound. On synthetic data, the authors compare the spread of estimates across repeated experiments and check how often the confidence intervals contain the true EOT value. In the synthetic experiment with a correct model, the method reduced the variance, and the confidence intervals had coverage close to 95%. When using a model with more parameters, the variance reduction almost disappeared. In the same experiments, the authors showed that when the assumptions about the model linking the distributions are violated, the estimate can systematically differ from the true value. On the OfficeHome dataset with real data, the method reduces the overall error when the target sample is small: its mean estimate becomes closer to the reference value, although its variance is higher than that of the standard empirical estimator.

Critical strengths

The authors determine how much the spread of EOT estimates can be reduced by using the relationship between the distributions, and prove that their method reaches this limit. Under the assumptions, no other regular estimator using the same information can have a lower asymptotic variance. This shows that the proposed method makes full use of the model linking the distributions to improve accuracy, without using additional data.

Critical weaknesses

Variance reduction is demonstrated in experiments where the density ratio model holds by construction. As the authors acknowledge, the paper does not establish that this benefit can also be obtained under natural distribution shifts. This leaves it unclear in which real-world settings the main theoretical advantage can be useful.

Experimental code is not provided with the paper or the supplementary material. Although the authors describe the experimental setup, the implementation and scripts needed to reproduce the tables and figures are missing.

Questions for authors

Could you give a specific example of real data with natural distribution shifts where the density ratio model is plausible and could reduce the variance of the EOT estimate? Are there experimental results for such a case?

3 comments

Natural shifts (W1/Q1). We compare Chinese A-share firms with positive or explicitly zero pre-tax amounts per share in their recorded annual cash dividend proposals. The features are return, volatility, and market capitalization at fiscal year-end, before disclosure (§6.3 and Appendix J.1–J.2). To study target scarcity, we set the sample sizes at 240 for the positive group and 20 or 40 for the zero group. We sample with replacement from each annual pool and use the same 600 draws per configuration for both estimators. RMSE is measured relative to the full-pool NP-EOT value (Appendix J.3, Appendix Table 17).

Across the 14 configurations in FY2018–FY2024, the 95% pointwise Monte Carlo intervals conditional on the annual pools lie below one for 12 DRM/NP variance ratios and all 14 RMSE ratios (main-text Table 3). For FY2024, the variance ratios are 0.637 [0.550, 0.746] and 0.618 [0.555, 0.692], reductions of 36.3% and 38.2% at target sizes 20 and 40. Both FY2023 variance intervals include one. We fit feature scaling on FY2018–FY2022 and hold it fixed for the later FY2023–FY2024 evaluation cohorts. These two pools share 4,645 firms (Appendix J.1), and Appendix Table 16 gives the group counts. The dividend comparison shows finite-sample variance reduction for a naturally observed grouping under the stated annual-pool design.

We use a linear tilt as a working model for the dividend groups. With G=1G=1 for an explicit zero proposal and λg=P(G=g)\lambda_g=P(G=g) before sampling, log odds a+β⊤za+\beta^\top z yield the ratio dF1/dF0=(λ0/λ1)exp⁡(a+β⊤z)dF_1/dF_0=(\lambda_0/\lambda_1)\exp(a+\beta^\top z) by Bayes' rule. The likelihood uses the fixed experimental sampling fractions. We obtain fitted scores by five-fold stratified cross-validation with class-balanced, ℓ2\ell_2-regularized logistic regression. The out-of-fold AUROC is 0.672 in FY2023 and 0.712 in FY2024. In both years over 99% of positive-group scores and all zero-group scores lie in [0.05, 0.95]. The diagnostic measures overlap on the fitted scale. Each EOT draw uses a separate unpenalized DRM fit (Appendix J.2). Theorem 1 concerns the asymptotic variance at a correctly specified pair under its stated conditions.

Code (W2). We prepared anonymous_code_reviewer.zip for the submission's Supplementary Material field. It contains the experiment scripts, frozen OfficeHome PCA features, and 8,400 pairs of saved dividend EOT estimates without firm-level features. From the extracted reproducibility/ directory, after installing requirements.txt, python scripts/verify_package.py checks package integrity and recomputes all 14 dividend summaries and intervals from the saved pairs. The command python experiments/dividend/export_tables.py --out outputs/dividend_tex regenerates the three dividend tables from the archived summaries. Small runs of the synthetic and both OfficeHome workflows passed. Refitting the dividend models and recomputing EOT require a separately prepared company-year feature snapshot that matches the archived checksum. See data/dividend/DATA_PROVENANCE.md for the authorized inputs and preparation steps and experiments/dividend/README.md for the replay commands. With the matching features, a small replay reproduced the FY2024 reference and 10 paired estimates.

I thank the authors for providing the source code and adding the dividend experiment. The variance reduction in FY2024 is substantial, and I also appreciate reporting the less clear results for FY2023. However, it is still unclear how well the linear tilt fits the data, and whether similar variance reductions can be expected in other settings.

My initial requests are mostly resolved. Given the additional concerns, I am still considering whether to keep my rating or raise it by one point.

Thank you for the follow-up. We chose a linear basis because it gives an interpretable model for how the odds of a zero-dividend proposal vary with return, volatility, and firm size. To check the fit, we added squared terms and pairwise interactions and compared the models on the same five folds within each year. Their out-of-fold log loss was nearly identical in FY2023 (both 0.567). In FY2024, the quadratic model improved on the linear model (0.560 vs. 0.570).

The 36–38% EOT variance reduction in FY2024 is specific to that cohort and sampling design. Both FY2023 variance-ratio intervals included one. Under correct specification, Theorem 1 explains why gains vary: their size depends on how much of the EOT potential contrast lies outside the chosen basis span. The theory already allows fixed nonlinear bases. A next step is to study how to select the basis from data while preserving valid inference for the EOT value.

Reviewer_4Ei8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work studies potential efficiency gains for entropic OT estimation when the two marginal distributions are related by an exponential tilt. The authors first introduce DRM-EOT, which couples the marginal distributions of the two samples through an estimated density ratio model (DRM). The new, weighted, marginal distribution can then be fed through the Sinkhorn algorithm to estimated the DRM optimal coupling. Theoretical analysis shows that this procedure has strictly lower variance than the standard nonparametric (NP) approach when the DRM is correct specified and joint potential contrast does not lie exactly in the linear span of Q. Numerical experiments confirm that, in the correctly specified setting this approach approach has lower variance than the standard approach. However, the results on the HomeOffice dataset appear to be mixed.

Critical strengths
  1. The theoretical analysis appears to be sound and has the potential to provide significant insight into nonparametric approaches for entropic OT estimation.
Critical weaknesses
  1. The manuscript is very difficult to read. Many sentences combine unrelated ideas, most of the mathematical notation is undefined or only implicitly defined, and there seems to be many missing citations given the long history of work on entropic OT and exponential tilting. For instance, equation (1) states that VNP−VDRM=ρ0ρ1E1(e2/ω)V_{NP} - V_{DRM} = \frac{\rho_0}{\rho_1} E_1(e^2/\omega), but V and "NP" are never defined. The manuscript states that ρ0\rho_0 and ρ1\rho_1 are sampling fractions, but not sampling fractions of what, and ω=ρ0+ρ1(dF1/dF0)\omega = \rho_0 + \rho_1 (dF_1/dF_0) but F1F_1 and F0F_0 are not defined. There are many instances like this which make it incredibly challenging to parse what is being presented.

  2. The method itself is a standard combination of profile likelihood estimation for two sample DRMs and the entropic OT algorithm. While the focus of this work is not on the method, but its asymptotic properties, my understanding is that the properties mostly state that 1. When the DRM is correct specified DRM-OT is more efficient (Theorem 1), 2. CLT gives a confidence interval for the OT plan (Theorem 2), and 3. under bounded misspecification, the method has bounded error (Section 5.2). These appear to be modest theoretical advances, using standard nonparametric statistical techniques, and their importance is not well discussed in the manuscript.

  3. The numerical experiments are also fairly limited. The synthetic example shows that there is a small variance improvement under a correctly specified model, as expected for a correctly specified parametric model. However, the real data example does not seem to show any tangible benefit for this method? Only on the n = 8000 setting with equal sampling ratio does Figure show a dip below the No Gain line.

Questions for authors
  1. Could the authors provide some insight into why the pooled empirical CDFs should work better than the raw empirical CDFs? The manuscript is not clear on this and it would require the reader to already be familiar with two sample density ratio methods.

  2. This approach should be validated on a wider range of empirical settings. The current OfficeHome experiments do not seem to show the efficiency gains that the method purports. Additionally, why does Q2 show no gain over NP?

  3. Please expand sections 4 and 5 to provide more insight into the theory being developed here? Its not clear what the implications for these equalities (Theorem 1) are or how the CLT result (Theorem 2) should be used. Moreover, section 4.1 is not very clear about what it is showing. The proof sketch of Theorem 1 should be moved to appendix and replaced with exposition on why it is important and what it means.

  4. Could you please more carefully define notation throughout the manuscript? Many equations and quantities are, at best, only implicitly defined through context, which makes reading and understanding this papers contributions quite difficult.

1 comment

W2/Q3. Efficiency and inference. We use established tools from empirical likelihood, Sinkhorn computation, and semiparametric projection. Our contribution is to quantify the EOT-relevant variation the joint DRM removes and prove attainment of the resulting bound. At positive limiting sampling fractions, Theorem 1 gives the exact asymptotic variance gain (ρ0/ρ1)E1(e2/ω)(\rho_0/\rho_1)E_1(e^2/\omega). The residual ee is the part of the joint potential contrast outside the tilt span in the weighted norm. The gain vanishes exactly when the contrast lies in the span. For two correctly specified nested bases at the same distribution pair and allocation, the larger span can shrink the residual norm and raise the bound (Appendix C).

Theorem 2 shows that the fitted estimator attains the bound. We derive the joint uniform process limit for the fitted marginals that Goldfeld et al.'s EOT delta method requires, identify the canonical gradient, and verify regularity along local DRM paths (Appendix D.3–D.6). Theorem 3 gives confidence intervals for the scalar EOT value. Inference for the optimal coupling is a separate problem. Under the assumptions of Theorem 2 and VDRM>0V_{\rm DRM}>0, the Wald interval uses standard error V^DRM/n\sqrt{\widehat V_{\rm DRM}/n}. The basic bootstrap interval resamples within each group and refits both the tilt and transport (Appendix E.1–E.2). The residual also gives the direction-specific mean shift under contiguous target departures (§5.2).

Additional result. Basis-dependent convergence rates. We added Proposition 1 (§5.3, Appendix K) for the same estimator at a fixed, correctly specified pair as the source size NN and target size mm both grow with m/N→0m/N\to0. If the first-order tilt derivative aa vanishes, positive source variance and nonzero tilt curvature give source and quadratic tilt terms of orders N−1/2N^{-1/2} and m−1m^{-1}, respectively. They have the same order at m≍Nm\asymp\sqrt N. In the uniform example, symmetry gives a=0a=0 for Q1=(1,x)⊤Q_1=(1,x)^\top. The added quadratic direction in Q2Q_2 restores a first-order term of order m−1/2m^{-1/2}. At the same distribution pair, these two correctly specified bases give different rates for the same estimation procedure.

Q1. Pooling. For r=dF1/dF0r=dF_1/dF_0, the weights 1/ωn1/\omega_n and r/ωnr/\omega_n recover the two marginals from Hn=ρ0nF0+ρ1nF1H_n=\rho_{0n}F_0+\rho_{1n}F_1, where ωn=ρ0n+ρ1nr\omega_n=\rho_{0n}+\rho_{1n}r. The fitted ratio lets both samples inform each marginal. The model links their fluctuations through the shared baseline and fitted tilt. Theorems 1 and 2 identify the EOT-relevant variation removed by the link and account for uncertainty in fitting the ratio (Appendix D.3–D.6).

W3/Q2. Controlled and natural-data results. In the synthetic design, the quadratic term in Q2Q_2 captures nearly all the even potential contrast. Quadrature gives a gain of 2.71×10−72.71\times10^{-7} on the VV scale. For Q1Q_1, the predicted and empirical reductions are 36.0% and 36.8% (main-text Table 1, Appendix G).

In exact-DRM OfficeHome, we prescribe the tilt on four fixed feature geometries and sample at (ρ0,ρ1)=(0.8,0.2)(\rho_0,\rho_1)=(0.8,0.2). The ESS ratio (ESS/NKN_K) measures target-mass concentration on the support. The source-to-target allocation stays 4:1 at both ESS settings. At n=16000n=16000, the empirical variance ratios are 0.950 [0.932, 0.969] and 0.971 [0.948, 0.995] at ESS ratios 0.5 and 0.8, with oracle ratios 0.952 and 0.929. Brackets give 95% pointwise Monte Carlo intervals. At n=8000n=8000 and ESS ratio 0.8, the interval includes one (main-text Table 2).

For the observed dividend proposal groups, 95% pointwise Monte Carlo intervals lie below one for 12 of 14 variance ratios and all 14 reference RMSE ratios, conditional on the annual pools. FY2024 variance reductions are 36.3% and 38.2% at target sizes 20 and 40. Both FY2023 variance intervals include one (§6.3, main-text Table 3). In natural OfficeHome, at nT=50n_T=50 DRM-EOT in the d=16d=16 baseline lowers mean absolute reference bias and reference RMSE, with a mean task-level variance ratio above one. The ratio of task-averaged reference RMSEs exceeds one at nT=400n_T=400 and 800800 (Appendix H.3 and H.9, Appendix Table 8).

W1/Q4. Exposition. We define the marginals, sample sizes, sampling fractions, NP and DRM bounds, and residual before Eq. (1). Section 2 relates our results to Goldfeld et al. (2024a, 2024b), Zhang and Chen (2023), and Chen and Fang (2019). Section 3.2 explains the sampling offset in Eq. (4), §4.1 gives the marginal derivative and weighted projection, and §4.2 interprets the gain. The proof outline is now in Appendix C.

Reviewer_rWktOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies the Entropic Optimal Transport (EOT) problem between the structurally related distributions. Their relationship is defined by the density ratio model with a fixed basis. Then the authors establish the exact variance-gain identity and propose the DRM-EOT approach which is proved to be semiparametrically efficient according to their theorem. The experiments are aimed at supporting the derived theoretical results.

Critical strengths

The paper results are theoretically justified.

Critical weaknesses

[W1] My main concerns are related to the limited clarity of the paper. The text is hard to follow. Many variables are introduced without explanations, e.g., VNP,  VDRMV_{NP},\;V_{DRM} (eq. 1), pgnp_{gn} (line 113), vector zz (lines 299-300), and transitions lack clarity, e.g., the intuition behind equation (4) might be clarified; the structure of experiments section is not clear since some subsections correspond to different experiments while some of them (6.2-6.4) - to data explanation and different aspects of the experiments on the same data. I kindly suggest the authors restructure this part and make the overall exposition of the paper more clear.

[W2] The efficiency-robustness trade-off analyzed in section 5.2 is based on the assumption that the distributions are shifted in the local way. In this case, the authors develop a bound on the mean shift. However, the assumed small perturbation of F1F_1 represents a particular case of possible perturbations and in my understanding does not support the general robustness of the methodology. As the authors state, they do not establish robustness to more general perturbations. The related experiment with controlled perturbation is provided in Appendix G, but it considers a rather simple setting. Thus, it raises concerns regarding the method’s applicability to real-world tasks where distributions’ relationship might be far from the assumed DRM.

[W3] In the OfficeHome experiment explanation (section 6.2), you define the basis Q using vectors zz which are not defined in this paragraph but if I correctly understand - correspond to the PCA representation. It means that the dimension of Q depends on the dimension of this PCA representation. Thus, the setup of the experiment in section 6.6 seems to be ambiguous for me. This experiment is aimed at analyzing the results for different dimensions dd, but it means that the dimension of the basis changes simultaneously. I am not sure that such a comparison is fair. I kindly ask the authors to comment on this point.

Questions for authors

[Q1] Is it possible to fix the experimental setup of section 6.6? See the weakness [W3]

1 comment

W3/Q1. Basis dimension at fixed transport geometry. The original sweep varied PCA geometry and the tilt basis together. The natural OfficeHome diagnostics (§6.2 and Appendix H.13) isolate basis dimension by fixing d=64d=64 and using the same normalized PCA features, transport cost, regularization, and complete-pool EOT reference. The DRM uses the first p=8,16,32,64p=8,16,32,64 coordinates at their existing scale in Qp(z)=(1,z1,…,zp)⊤Q_p(z)=(1,z_1,\ldots,z_p)^\top. The four fits share draws and NP-EOT estimates across 12 directed tasks, with nS=1000n_S=1000, nT=50,200,800n_T=50,200,800, and 100 repetitions per cell, sampling without replacement within each group. We retain every draw and use the corresponding NP-EOT estimate for each of the 53 failed fits.

At nT=50n_T=50, the mean task-level DRM/NP RMSE ratios are 1.470, 0.830, 0.617, and 0.725 for the four bases. The extreme-tilt frequencies are 0%, 0%, 10.75%, and 60.00%. Mean variance ratios exceed one throughout the grid, and mean RMSE ratios exceed one for all four bases at nT=200n_T=200 and 800800 (Appendix Tables 13–14). Appendix Table 15 gives the task-level results. In this natural comparison, changing the basis can affect both approximation error and fitting stability. The oracle calculation at the end of Appendix H.13 checks the variance ordering under exact specification. Every nested basis contains the true tilt, and the oracle variance ratio is nondecreasing in pp, as predicted by Theorem 1.

W1. Notation and organization. Before Eq. (1), we define the marginals, sample sizes, sampling fractions, projection residual, and efficiency bounds for the nonparametric (NP) and DRM models. Section 3.2 explains the logistic form in Eq. (4) and how the fitted ratio reweights the pooled observations for both marginals. Section 6.2 defines zz by projecting frozen ResNet-50 penultimate-layer features onto the first dd principal components and normalizing. PCA is fitted on the three non-target domains. Section 6 presents the controlled synthetic study, exact-model OfficeHome validation and natural-shift diagnostics, and the dividend proposal study.

W2. Misspecification. Equations (12)–(13) in §5.2 give the local bias coefficient and its bound along any fixed, bounded, mean-zero target direction at the n−1/2n^{-1/2} scale. The calculation starts from a correctly specified pair, holds the source marginal fixed, and assumes positive limiting sampling fractions. The same residual ee enters the variance gain and the local bias coefficient Bδ=−ρ0E1(eδ/ω)B_\delta=-\rho_0E_1(e\delta/\omega). Appendix G varies the perturbation magnitude along one specified quadratic direction. Fixed or nonlocal misspecification calls for a separate analysis. For the natural-shift experiments, we compare variance and reference RMSE with NP-EOT. Both FY2024 pointwise Monte Carlo variance-ratio intervals lie below one, and both FY2023 intervals include one (§6.3, main-text Table 3). In the natural OfficeHome baseline at d=16d=16, smaller reference bias accompanies lower RMSE at small target sizes despite a mean task-level variance ratio above one (Appendix H.3 and H.9, Appendix Table 8).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.