acceptodds
Under review as a conference paper at ICLR 2027

Forward Tangent-Centered Sinkhorn DRO: Exact Nominal Recovery from Optimal Transport to KL

17 views

Abstract

# ICLR Abstract Fixed-reference entropy-regularized transport yields smooth distributionally robust optimization (DRO) objectives, but its value as a function of the adversarial marginal is generally minimized at a kernel-smoothed law rather than at the nominal distribution. Consequently, the smallest raw ambiguity set need not recover nominal risk. Building on prior information-geometric plan-KL orientations, we study the forward tangent-centered discrepancy obtained by subtracting the value and first variation at the nominal law. This is the unique affine recentering, up to the simplex gauge; for symmetric costs, its plan-KL representation coincides with the forward candidate noted by Amari et al. (2018). The representation shows that the discrepancy dominates marginal KL, is uniquely zero at the nominal law, and is strongly convex in total variation. The resulting DRO problem retains a scalar log-sum-exp dual and closed-form conditional adversarial weights. On finite supports with symmetric zero-diagonal costs, the discrepancy converges uniformly to the unregularized transport cost at low temperature; for any finite cost, its rescaled value converges uniformly to marginal KL at high temperature. We derive its exact local Hessian, an explicit finite-sample containment radius, and a generalized chi-square calibration for multinomial data. Controlled simulations verify nominal centering, both limiting regimes, and practical plug-in coverage.

Primary area:
optimization
Keywords:
Distributionally Robust Optimization, Optimal Transport, Entropy-Regularized Optimal Transport, Sinkhorn Divergences, Convex Optimization, Information Geometry, Statistical Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_chH6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the issue of nominal center drift in Distributionally Robust Optimization (DRO) with a fixed reference distribution. The authors propose a divergence measure centered on the tangent, which recovers the nominal distribution as its unique minimizer. They establish a representation based on KL-divergence, derive the dual formulation for DRO with scalar, and examine the limiting behavior as the regularization temperature varies. Furthermore, the paper provides a local Hessian characterization and results on statistical calibration, and conduct numerical experiments to verify the characterization.

Critical strengths
  1. This paper identifies a clear issue with Sinkhorn DRO based on a fixed reference measure. The concept of centering around a tangent plane is both simple and insightful.

2 .The theoretical analysis is quite comprehensive: the scalar dual formulation, the temperature limit, and the results concerning the local Hessian fit together well.

Critical weaknesses
  1. The main concern is the novelty of this papaer. The forward projection KL-divergence term already appears in Amari et al. (2018), while "tangent centering" is a standard construct within the framework of Bregman divergences. The associated Sinkhorn-DRO duality and the corresponding statistical results have also been well studied. While this paper does a good job of synthesizing these components, I do not believe it represents a sufficiently substantial new contribution.

  2. The validation seems impractical. No actual DRO learning task is presented to demonstrate why this approach is superior to existing alternatives. Furthermore, the reported experiments do not satisfy the conditions of finite samples specified in Theorem 5.2.

Questions for authors
  1. What is the main theoretical contribution beyond the plan-KL geometry already discussed by Amari et al. (2018)?

  2. Can the authors give a concrete example where tangent-centered DRO leads to better decisions than existing Sinkhorn-DRO or KL-DRO?

  3. How useful is the finite-sample radius in practice, given that its sufficient condition is not met in Table 2?

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

The authors study a discrepancy DγTCD_\gamma^{\mathrm{TC}} for distributionally robust optimization (DRO) obtained by subtracting an affine term from the entropic optimal transport (eOT) to make it centered at the nominal measure pp. The paper establishes that the considered affine correction is unique to make the resulting discrepancy minimal at pp. It also shows that the discrepancy is proportional to KL(Πγq,p∣∣Πγp,p)\mathrm{KL}(\Pi_\gamma^{q,p}||\Pi_\gamma^{p,p}), where qq is the adversarial measure, and Πγq,p,Πγp,p\Pi_\gamma^{q,p}, \Pi_\gamma^{p,p} are eOT plans. Thus, for symmetric costs, DγTCD_\gamma^{\mathrm{TC}} coincides with the divergence proposed by Amari et al.; the authors contribute by studying its properties and application to DRO. It is shown that DγTCD_\gamma^{\mathrm{TC}} is convex in the adversarial measure and DγTC(q∣∣p)≥γKL(q∣∣p)D_\gamma^{\mathrm{TC}}(q||p) \geq \gamma\mathrm{KL}(q||p).

The authors establish that adversarial risk has a fairly simple form and involves an infimum over an auxiliary scalar variable. Moreover, the dual representation of the DRO objective is shown to be jointly convex in the model parameters and the auxiliary variable if individual losses are convex. As the temperature grows, the discrepancy normalized by the temperature converges to KL(q∣∣p)\mathrm{KL}(q||p). If the cost matrix is symmetric, non-negative, and has zero diagonal, then the discrepancy converges to the unregularized OT as the temperature goes to zero. Limiting behavior of the respective DRO problems follows. Finally, the authors derive an exact formula for the Hessian at pp and specify finite sample and asymptotic radii, such that the respective uncertainty set contains the true distribution with high probability under appropriate assumptions. The theory is illustrated by experiments on synthetic data.

Critical strengths

The studied divergence is shown to possess favorable mathematical properties. The paper seems mathematically sound.

Critical weaknesses
  • It is not clear if the studied discrepancy is practically useful. I suggest that the authors consider real DRO problems and investigate if DγTCD_\gamma^{\mathrm{TC}} brings practical gains compared to other discrepancies.
  • The mathematical techniques used in the paper are fairly standard.
Questions for authors
  1. Please improve readability of the paper.
    • To avoid ambiguity, most terms should be introduced before the first use, e.g., reference measure, reference matrix, nominal law, gauge, dual candidate, etc.
    • Some sentences should be made more easily understandable, e.g., "using the forward Bregman orientation dictated by the adversarial marginal" (line 50).
    • Upon seeing Eq.(2) for the first time, a reader should wonder if this is a definition or a property, and what the superscript TC means.
  2. Can you please add a detailed comparison of DγTCD_\gamma^{\mathrm{TC}} with the "self-plan-first" construction of Amari et al. in the context of DRO, and in particular reflect the differences in Table 1?
  3. Additional minor remarks:
    • Line 99: "TγT_\gamma is defined by (1). This is entropy relative to counting measure, up to a constant" - this part is formulated as if you're claiming that TγT_\gamma is entropy relative to counting measure, which is incorrect In general.
    • Line 108: "minimize Tγ(q,p)T_\gamma(q,p) over Π\Pi" - strictly speaking, TγT_\gamma doesn't depend on Π\Pi.
Reviewer_zJoqOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies centering entropic optimal transport for ambiguity sets in distributionally robust optimization (DRO) on finite supports. Subtracting the value and first variation of the transport functional at the nominal distribution removes the minimizer’s displacement. The resulting discrepancy admits a representation as the KL divergence between the optimal transport plan and the nominal self-coupling. The authors derive a scalar dual formulation, uniform limits to discrete OT and, after rescaling, to KL divergence, a local Hessian, and statistically calibrated radii for multinomial data. Numerical experiments examine centering, the limiting regimes, and the coverage of the asymptotic plug-in radius

Critical strengths
  1. The paper gives a coherent solution to the displacement of the center caused by entropic regularization. The proposed correction preserves convexity and a tractable dual formulation, and recovers the nominal risk exactly at zero radius.
  2. The main proofs appear correct. Useful results include the explicit relationship between plan-level and marginal KL divergences, uniform convergence over the adversarial marginal in the limiting regimes, and a compact expression for the Hessian in terms of a conditional covariance matrix. Uniform convergence justifies limits of ambiguity sets and robust risks, not just pointwise limits for a fixed pair of distributions.
  3. The authors appropriately acknowledge connections to prior work and explicitly restrict their guarantees to a fixed finite support and a given cost function. They also separate the finite-sample guarantee from the asymptotic calibration evaluated experimentally.
Critical weaknesses
  1. The novelty beyond prior work is limited. The forward plan-KL construction appears in Amari et al. (2018, Eq. 50), as the authors note. Uniqueness of the affine centering follows from the first-order optimality condition. The scalar dual formulation and exponential conditional weights follow by applying the existing Sinkhorn-DRO duality of Wang et al. (2025) to the corrected cost. The mechanism behind the generalized (\chi^2) limit is known from Bigot et al. (2019). Moreover, the matrix [ M_{\gamma,p}=\operatorname{Diag}(p)-P\operatorname{Diag}(p)^{-1}P^\top ] is related through a Schur complement to the block covariance matrix in Amari et al. (2018, Theorem 5). The main value of the paper lies in integrating existing constructions, providing explicit bounds, and specializing them to DRO. The uniform low-temperature limit is the most substantive contribution, but the overall advance over prior work is not sufficiently convincing.
  2. The finite-sample guarantee can be uninformative. For the experimental parameters (m=5), (a=1/2), (\delta=0.05), and (\underline p=0.08), the condition in Theorem 5.2 first holds at (n=8386), above every sample size in Table 2. At (n=10000) and (\gamma=0.2), Eq. (21) gives (\rho^{FS}\approx73.8). Lemma D.1 implies [ \max_q D_\gamma^{TC}(q|\widetilde p_n) \le M_C+\gamma\log\frac{m}{\min_j\widetilde p_{n,j}} \le3.303 ] for every sample realization under the smoothing scheme used in the paper. The guaranteed radius covers the entire probability simplex, so the robust objective reduces to the maximum loss across scenarios. This does not contradict the theorem, but the guarantee is uninformative here. The paper should identify parameter ranges where the guarantee restricts the ambiguity set nontrivially.
  3. The experiments do not show that centering helps solve a DRO problem. All examples use three or five support points and check algebraic properties or coverage. No experiment optimizes a decision or model parameter (\theta) and evaluates out-of-sample performance. Comparisons with nominal optimization, the original Sinkhorn-DRO formulation, KL-DRO, and unregularized Wasserstein-DRO are missing. It is unclear whether removing the center’s displacement improves decision quality or whether the transport geometry justifies the extra computation. Large-scale benchmarks are not required for a theoretical paper, but given the moderate novelty, one convincing example of an advantage would strengthen the contribution. Radii should use comparable coverage levels or a common validation protocol.
  4. The statistical calibration evaluation is too narrow. The experiments use one distribution on five support points, without rare categories, larger supports, or lower temperatures. At (\gamma=5), the local geometry at the chosen distribution is nearly identical to KL: the largest nonzero eigenvalue of (\Sigma^{1/2}H_{\gamma,p^\star}\Sigma^{1/2}/\gamma) is approximately (1.00184), while all such eigenvalues equal one for KL. This regime barely tests effects specific to transport geometry. Furthermore, at (n=50,\gamma=5), coverage is (93.55%) rather than (95%). This does not invalidate the asymptotic theorem, but it requires more cautious interpretation than noting the estimate is within approximately three standard errors of the target. Using a shared bank of 5,000 Gaussian draws is valid, but the reported standard errors ignore variability across realizations of that bank. Quantile approximation error should be separated from the error of the asymptotic approximation.
Questions for authors
  1. Which result is the principal theoretical advance over the forward candidate of Amari et al., existing Sinkhorn-DRO duality, and the second-order delta method? What technical challenges are needed to establish it?
  2. Can the finite-sample radius be sharpened, or can you identify a practically relevant regime where it does not include the entire simplex while the geometry remains substantially different from KL?
  3. Can you provide a task where tangent centering improves decision quality relative to the original Sinkhorn-DRO formulation and KL-DRO under comparable radius calibration?
  4. Does coverage remain adequate at lower (\gamma), with rare categories, and as (m) increases? Does the undercoverage persist when quantiles are computed more accurately?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.