acceptodds
Under review as a conference paper at ICLR 2027

Extending Rademacher Complexity-Based Generalization Theory to Domain Generalization

Abstract

Under the independent and identically distributed (i.i.d.) assumption, Rademacher complexity-based statistical learning theory characterizes the model's generalization through a Rademacher complexity term and its high-probability confidence guarantee. However, this framework becomes inadequate in the domain generalization scenario where distribution shifts commonly occur, precisely because distribution shift itself gives rise to an additional, distribution-induced model complexity, whereas classical Rademacher complexity is conditioned solely on training samples and thus fails to capture the complexity variations induced by distribution shifts. To overcome this limitation, we propose to treat domain distributions as the fundamental objects for defining distribution-level complexity. Specifically, we introduce a risk functional class on the Wasserstein space and quantify the model complexity with respect to domain distributions via its Rademacher term. Furthermore, by exploiting the curvature structure of the Wasserstein space, we modify the McDiarmid-type concentration inequality to obtain a Wasserstein-geometric confidence guarantee for the proposed complexity measure. Based on this complexity measure, we derive a domain generalization error bound that separates a distribution-level Rademacher complexity term, which captures the complexity caused by distribution shift, and that establishes a connection between the curvature of the sample space and domain generalization performance through its confidence term. This reveals a new principle of domain generalization: learning representation space with higher curvature facilitates better cross-domain generalization. Finally, we experimentally validate this conclusion.

Primary area:
learning theory
Keywords:
Domain Generalization, Machine Learining Theorty, Wasserstein space, Rademacher complexity

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_UbmEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends Rademacher complexity-based generalization theory to domain generalization by treating source domains as random distributions and defining a distribution-level Rademacher complexity on the Wasserstein space.

Critical strengths
  1. The idea of connecting optimal transport geometry and curvature to DG generalization is new and interesting, and may inspire further work.
  2. The proofs are written out in full in the appendix and are mostly easy to follow.
  3. The authors try to validate the theory with experiments on three standard DG benchmarks, and openly discuss the cases where results do not match (PACS Photo, Office-Home Clipart)
Critical weaknesses
  1. The hierarchical setting, where domains are drawn from a meta-distribution and then samples from each domain, and the resulting domain-level and sample-level complexity terms, are well established in learning-to-learn and DG theory [1,2].
  2. In classification, Y is a finite set, so Z = X × Y is not a length space and CD(K,∞) with K > 0 cannot hold. The paper does not explain how Theorem 3 applies in its own experiments.
  3. The DG error values in (Figures 1 and 2). With a linear classifier, a smaller radius gives smaller logits and therefore smaller loss gaps, regardless of generalization. Since the source and target accuracy are not reported, there is a speculation that the smaller gap may come from underfitting.
Questions for authors
  1. How does a union of disjoint spheres satisfy CD(K,∞) with K > 0? What is the actual K of the latent measure in your experiments?
  2. How do you handle discrete labels in Z?
  3. How do you account for the fact that B/√K is invariant to rescaling the metric? Under what condition does higher curvature truly reduce the bound?
  4. How does your two-level bound compare to the meta-learning bounds of [1] and [2]?
  5. What exactly is the “DG error” in Figures 1 and 2? Can you report source and target accuracy for each radius?

REFS:

[1] Maurer et al, The Benefit of Multitask Representation Learning

[2] Blanchard et al, Domain Generalization by Marginal Transfer Learning

4 comments

Weaknesses 1 and Question 4: We thank the reviewer for pointing us to these important references. We agree that the hierarchical setting (drawing domains from a meta-distribution and then samples from each domain) is well established in learning-to-learn and domain generalization theory, including Maurer et al. (2016) and Blanchard et al. (2021). But we do not claim this hierarchical framework as an independent contribution; in Section 2.2 we introduce it following [3] (Gouk et al. (2022)) for completeness.

Regarding the comparison with the bounds in [1] and [2], we would like to clarify the relationship. We explicitly restrict the domain space to a CC-regular subset of the 2-Wasserstein space, whereas [1] and [2] do not impose this Wasserstein restriction. This directly lead to our modification of McDiarmid's inequality and to the curvature-dependent confidence term, which is not present in the bounds of [1] and [2]. We therefore view our bound as complementary to [1] and [2]: it does not aim to replace their more general or potentially tighter bounds, but provides a geometric perspective that connects the curvature of the representation space to domain generalization.

Weaknesses 2 and Question 2: We thank the reviewer for this insightful comment, and we fully agree with the criticism. In classification, YY is a finite set, so Z=X×YZ=X\times Y is not a length space. We propose the following revision. Following [4], [6], and [7], we assume that each domain has a labeling function y=fD(x)y=f_D(x), and adopt the covariate shift assumption (common in domain generalization theory) fD1(x)=fD2(x)f_{D_1}(x)=f_{D_2}(x) for all D1,D2∈DCD_1,D_2\in\mathbb{D}^C. Under this assumption, we define the domain distributions only on the representation space ZZ, i.e., we have changed the meaning of the symbol ZZ from the product space X×YX \times Y to the representation space. The risk functional becomes ϕℓ,h(D)=Ex∼D[ℓ(h(x),f(x))],\phi_{\ell,h}(D)=E_{x\sim D}\big[\ell(h(x),f(x))\big], where ff is the shared labeling function. See the beginning of Section 2.2 in the revised manuscript for details.

Weaknesses 3 and Question 5: We thank the reviewer for these detailed concerns. We agree that using a single hypersphere may lead to underfitting, and that the mixture-of-hyperspheres construction was introduced to mitigate this issue but may not eliminate it entirely; indeed, variations in training accuracy are observed in practice. Therefore, we can only verify the relationship between curvature and generalization error by showing how this empirical generalization gap varies, which does introduce a limitation in experimental rigor. We will add explicit discussion of this experimental limitation in the revised version (Appendix G.4).

Question 1: In our setting, the representation space is a Riemannian manifold, the base manifold is a union of hyperspheres equipped with the round metric. For a standard sphere of radius RR, the Ricci tensor is n−1R2g,\frac{n-1}{R^2}g, where nn is its dimension, RR is the radius, and gg is the Riemannian metric tensor. In our experiments, we set KK to the sectional curvature 1/R21/R^2. Then, by the equivalence theorem given in Theorem 7.3(b) ((1) and (3)) of [5], one can directly verify that our experimental setting satisfies the CD(K,∞)\mathrm{CD}(K,\infty) condition.

Question 3: We thank the reviewer for this insightful question. We agree that the curvature-dependent term indeed couples with the Lipschitz constant qzq_z and the domain regularity constant CC, and these quantities need not remain fixed when the representation changes.

To address the coupling with qzq_z, we will remove the Lipschitz assumption on the loss and replace the Kantorovich--Rubinstein duality step in the proof of Theorem 3 with the domain adaptation inequality in Lemma 1 of [1] (Redko et al. (2019)), inheriting its assumptions. This modification eliminates the explicit qzq_z factor from the bound. We will add an appendix summarizing the relevant lemma (Appendix D in the revised paper) and give the revised derivation of Theorem 3 (Eqn. (31) in the revised paper).

To address the coupling with CC, we will no longer treat CC as a fixed constant but keep it as an explicit, representation-dependent factor in the bound, around which we build an explicit invariance interpretation. See the revised version of the paper for details (Definition 4, Eqn. (21), (23), (24)). Guided by this invariance interpretation, we revise the claim about curvature. The original statement that "higher curvature leads to better domain generalization'' is replaced with "domain generalization is jointly affected by the curvature of the representation space and its degree of cross-distribution invariance.'' This statement directly connects our bound to invariant representation learning for domain generalization.

[1] Maurer et al, The Benefit of Multitask Representation Learning.

[2] Blanchard et al, Domain Generalization by Marginal Transfer Learning.

[3] Gouk et al, On the limitations of general purpose domain generalisation methods.

[4] Redko et al, Theoretical analysis of domain adaptation with optimal transport.

[5] Lott et al, Ricci curvature for metric-measure spaces via optimal transport.

[6] Albuquerque et al, Generalizing to unseen domains via distribution matching.

[7] Zhao et al, On learning invariant representations for domain adaptation.

Thank you authors for the rebuttal. Most of my concerns, particularly regarding (W2,Q2), (W3,Q5), and Q3 are still unaddressed. Infact, since that the authors acknowledge for the change the assumption (in response to W2,Q2), and removing assumptions from Theorem 3, how it has affected the error bound is unanswered. Furthermore, the speculation of underfitting due to use of small radius (of single hypersphere) and the mitigation through use of mixture of hypersphsre is not quantified in the response to W3,Q5. Due to these reasons, i choose to stick with my inital rating of weak reject (2).

We thank the reviewer for the response, and we would still like to provide some explanations regarding (W2, Q2), Q3, and our revisions to these issues. First, we have submitted a revised version of the manuscript, and in each response below we indicate where the corresponding revision was made. For further details, please refer to the revised version.

  1. Our revision for (W2, Q2) indeed addresses this concern of yours: following Redko et al. (2017), Albuquerque et al. (2019), and Zhao et al. (2019), we assume that each domain has a labeling function y=fD(x)y=f_D(x) and adopt the covariate shift assumption fD1(x)=fD2(x)f_{D_1}(x)=f_{D_2}(x) for all D1,D2∈DCD_1,D_2\in\mathbb{D}^C. We then define domain distributions only on the representation space ZZ, and the risk functional becomes ϕℓ,h(D)=Ez∼D[ℓ(h(z),f(z))],\phi_{\ell,h}(D)=E_{z\sim D}\big[\ell(h(z),f(z))\big], where ff is the shared labeling function. The same replacement is made in all subsequent formulas, and the derivations are modified accordingly. This revision indeed removes the discrete space YY from the definition of the underlying sample space, so that ZZ represents only the representation space (which is a standard length space). The effect on the bound is as follows. The mathematical form of the final bound is unchanged, but its interpretation and validity are materially different: K(Z)K(Z) is now the Ricci curvature lower bound of the representation space ZZ, the domain space is P2(Z)P_2(Z), and all Wasserstein-geometric steps (including the CD(K,∞)\mathrm{CD}(K, \infty) condition) now act on ZZ, which is a length space. In the original manuscript, the curvature assumption was imposed on X×YX \times Y, where it cannot hold for finite label sets. In the revision, the assumptions and the conclusion are consistent on the same space ZZ.

  2. For Q3, our understanding of your question is as follows: changing the curvature causes the representation space to change, making BB not a constant but a variable that depends on the representation space, so the curvature couples with qzq_z and CC in the original manuscript (B=2qz2C−2B=2q_z\sqrt{2C-2}). To address the coupling with qzq_z, we remove the Lipschitz assumption on the loss and replace the Kantorovich–Rubinstein duality step in the proof of Theorem 3 with the domain adaptation inequality in Lemma 1 of Redko et al. (2019): ∣ED1(h,h′)−ED2(h,h′)∣≤W1(D1,D2)|\mathcal{E} _ {D _ 1}(h,h^{\prime}) - \mathcal{E} _ {D _ 2}(h,h^{\prime})| \leq W _ {1}(D _ 1, D _ 2) This revision directly replaced the content in the original manuscript (the content below Eq. (30) and above Eq. (31)) with Eq. (31) in the revised manuscript, thereby eliminating qzq_z. As for the bound itself, this only eliminates qzq_z and has no other impact. To address the coupling with CC, we no longer treat CC as a fixed constant. Instead, we keep it as an explicit representation-dependent factor and use it to define the invariance measure Inv(Z)=1Cmin⁡−1,\mathrm{Inv}(Z)=\frac{1}{\sqrt{C_{\min}-1}}, The final bound resulting from these two revisions is as follows: sup⁡h∈HErrD(h)≤2R^D1,⋯ ,Dm(Φℓ,H)+2qm∑i=1mR^ni(H)+O((ln⁡(2/δ)n)1/2)+O((ln⁡(2/δ)m⋅Inv⁡(Z)2⋅K(Z))1/2).\sup_{h \in \mathcal{H}} \mathrm{Err} _ D(h) \leq 2 \hat{\mathcal{R}} _ {D _ 1, \cdots, D _ m}(\Phi _ {\ell, \mathcal{H}}) + \frac{2q}{m}\sum_{i=1}^m \hat{\mathcal{R}}^i _ n(\mathcal{H}) + \mathcal{O}\Big(\big(\frac{\ln(2/\delta)}{n}\big)^{1/2}\Big) + \mathcal{O}\Big( \big(\frac{\ln(2/\delta)}{m\cdot\operatorname{Inv}(\mathbb{Z})^2\cdot K(\mathbb{Z})}\big)^{1/2} \Big). Please refer to Theorem 4 in the revised version for details. Therefore, we believe that, with these solutions, we can answer Q3: higher curvature reduces the bound under the condition that the invariance Inv(Z)\mathrm{Inv}(Z) does not degrade, i.e., CC does not increase. Equivalently, simultaneously learning invariant representations and increasing the curvature of the representation space reduces the domain generalization gap. If the metric is rescaled, KK and CC change together, and one cannot claim a benefit from curvature alone. Our revised claim is therefore that domain generalization is jointly affected by the curvature of the representation space and its degree of cross-distribution invariance, rather than that higher curvature always helps.

we hope this response addresses the reviewer's concerns. We welcome any further questions or suggestions. If the concerns are resolved, we would be grateful if the reviewer considers updating their score accordingly.

Reviewer_FF9qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies domain generalization under a hierarchical sampling model, where source domains are sampled from a meta-distribution and training examples are subsequently sampled within each domain. It defines a class of domain-risk functionals and analyzes its Rademacher complexity. Under a positive curvature condition CD(K,∞)CD(K,\infty), a uniform Lipschitz assumption, and a bounded density-ratio condition, the paper derives a curvature-dependent bounded-difference constant and a generalization bound separating domain-level and within-domain contributions. Experiments on PACS, Office-Home, and Digits-DG investigate the effects of spherical representation radius and source-domain count.

Critical strengths
  1. The hierarchical sampling perspective is appropriate. Separating uncertainty from sampling domains and uncertainty from sampling examples within domains addresses an important feature of domain generalization that a pooled-sample analysis can obscure.
  2. The geometric bounded-difference argument provides an interpretable connection. Combining loss Lipschitzness, transport–entropy inequalities, and restrictions on domain densities gives an explicit relationship between domain-risk variation and geometric assumptions. This is the most promising technical aspect of the paper.
Critical weaknesses
  1. The theoretical results provide limited support for the central claim about curvature and domain generalization. Curvature enters the bound through a confidence term, while the complete guarantee also depends on domain-level complexity, within-domain complexity, loss Lipschitzness, and the restriction on domain distributions. The paper does not characterize how these quantities change together when learning representations. Consequently, the bound does not establish when increasing representation curvature improves the overall guarantee or unseen-domain performance. The experiments show an association between representation radius and generalization, but do not distinguish the proposed geometric explanation from changes in feature scale, stochasticity, or optimization. The resulting insight is therefore less conclusive than the paper’s main claim suggests.

  2. The proposed Rademacher complexity is not shown to provide sufficient explanatory or algorithmic value for DG. Separating domain-level and sample-level uncertainty is reasonable, but the paper does not demonstrate whether the proposed complexity yields informative bounds for the studied models, explains performance differences between DG methods, or guides useful algorithmic choices. The experiments examine representation radius and source-domain count without measuring or controlling the complexity term. As a result, they provide limited evidence that this quantity helps explain DG beyond the general observation that training on finitely many domains introduces additional uncertainty. A clearer connection between the complexity measure, observed generalization, and actionable learning principles is needed to substantiate the contribution.

Questions for authors
  1. Under what conditions does increasing representation curvature improve the complete generalization bound? How does the analysis account for simultaneous changes in the complexity terms and loss Lipschitzness?

  2. What distinctive prediction does the framework make about when a DG method will succeed or fail, beyond the expected benefit of observing more source domains? Is this prediction supported by the experiments?

  3. Can the proposed domain-level Rademacher complexity explain performance differences between representations or DG algorithms, or motivate a concrete algorithmic choice?

2 comments

Weaknesses 1 and Question 1: We thank the reviewer for this insightful comment. We agree that there is coupling between curvature term and each of the complexity, the loss-Lipschitz constant, and the domain-regularity constant CC. Below we clarify each point or provide a corresponding solution one by one:

  1. First, the Rademacher complexity term depends on the representation's complexity and therefore will inevitably couple with the curvature of representation space. However, any theory that analysis in a representation space faces this issue. For example, invariant representation learning theory based on distribution discrepancy measures ([3]) also changes the representation space when learning invariant representations, thereby causing changes in complexity. We therefore believe that this coupling is theoretically difficult to avoid.

  2. To address the coupling with qzq_z, we will remove the Lipschitz assumption on the loss and replace the Kantorovich--Rubinstein duality step in the proof of Theorem 3 with the domain adaptation inequality in Lemma 1 of [1] (Redko et al. (2019)), inheriting its assumptions. This modification eliminates the explicit qzq_z factor from the bound. We will add an appendix summarizing the relevant lemma (Appendix D in the revised paper) and give the revised derivation of Theorem 3 (Eqn. (31) in the revised paper).

  3. I also candidly acknowledge that the curvature couples with the domain-regularity constant CC, a point that was not mentioned in your comment. To address this, we will no longer treat CC as a fixed constant but keep it as an explicit, representation-dependent factor in the bound, around which we build an explicit invariance interpretation. See the revised version of the paper for details (Definition 4, Eqn. (21), (23), (24)). Guided by this invariance interpretation, we revise the claim about curvature. The original statement that "higher curvature leads to better domain generalization'' is replaced with "domain generalization is jointly affected by the curvature of the representation space and its degree of cross-distribution invariance.'' This statement directly connects our bound to invariant representation learning for domain generalization.

Weaknesses 2 and Question 2, 3: We thank the reviewer for this critical comment. We agree that our manuscript did not sufficiently connect the complexity measure to observed generalization or to actionable learning principles, and that the experiments did not measure or control the complexity terms.

First, we would like to clarify the role of the distribution-level Rademacher complexity. By itself, it provides a theoretical characterization of the excess uncertainty induced by training on a finite number of source domains. Our main new interpretation comes from combining this term with the curvature-dependent confidence term: the bound separates domain-level uncertainty from sample-level uncertainty and shows that the confidence term depends on the Ricci curvature of the representation space.

Regarding algorithmic value, we will clarify that the complexity measure provides an explanatory principle rather than a directly computable regularizer. The domain-level complexity penalizes the model's ability to fit domain-specific information; since domain-invariant information is consistent across domains, a representation that relies more on domain-specific information incurs a larger penalty. This connects our bound to invariant representation learning: methods such as domain-adversarial training, IRM, and domain diversification can be viewed as reducing the domain-level complexity penalty by suppressing domain-specific information or increasing the diversity of source-domain draws.

[1] Redko, Ievgen and Habrard, Amaury and Sebban, Marc. Theoretical analysis of domain adaptation with optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 737--753 (2017).

[2] John Lott and C´edric Villani. Ricci curvature for metric-measure spaces via optimal transport. Annals of Mathematics, pp. 903–991, (2009).

[3] Zhao, Han and Des Combes, Remi Tachet and Zhang, Kun and Gordon, Geoffrey. On learning invariant representations for domain adaptation. International conference on machine learning, pp. 7523--7532 (2019).

Thank you for the detailed response and clarifications. I appreciate the authors' efforts to refine the theoretical claims and acknowledge the coupling between curvature and other terms in the generalization bound. However, the response still does not establish when curvature improves the complete bound, nor does it provide theoretical or empirical evidence that the proposed Rademacher complexity explains differences in DG performance or guides algorithmic choices. Therefore, my main concerns remain, and I will maintain my original rating.

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

The paper studies domain generalization through Rademacher complexity defined over domain-risk functions. It derives a curvature-dependent concentration bound and argues that higher-curvature representations improve generalization. Experiments on PACS, Office-Home, and Digits-DG examine changes in representation radius and source-domain count.

Critical strengths

The connection between representation geometry and domain generalization is worth studying. The paper attempts to connect its analysis with experiments on three established benchmarks.

Critical weaknesses
  1. The paper incorrectly states a result from prior work, and the novelty is overstated.

    Theorem 2 and the explanation below Eq. (11) misrepresent Gouk et al.’s result [1]. The original theorem uses a sample set containing one example from each source domain. This paper instead describes samples from a single source domain. Gouk et al. also already define domain-level Rademacher complexity in their Eq. (7). Treating domains as the objects of a Rademacher analysis is therefore not a new contribution. The paper does not adequately distinguish its proposed complexity measure from this existing construction.

  2. The proof of the main bound contains an invalid step.

    The symmetrization argument in Eq. (48) exchanges an absolute value and an expectation as an equality, then drops absolute values when introducing Rademacher signs. The resulting inequality is false in general. Under the definition in Eq. (19), a singleton function class has zero Rademacher complexity, while its expected absolute deviation can be positive. The proof therefore does not establish Theorem 4 as written.

  3. The claimed benefit of higher curvature does not follow from the bound.

    The curvature-dependent term also depends on the Lipschitz constant qzq_z and the domain regularity constant CC. These quantities need not remain fixed when the representation changes. Even a simple rescaling can increase KK without reducing qz/Kq_z/\sqrt{K}. Moreover, Eq. (10) concerns average risk under the same meta-distribution, rather than performance on arbitrary or worst-case target domains. The claim that higher curvature improves cross-domain generalization is broader than what the analysis establishes.

  4. The experimental representation does not establish the geometry required by the theory.

    Equation (56) combines sphere centers through soft assignment weights. As written, the resulting features do not generally lie on the claimed union of fixed spheres. The paper also does not establish the required curvature-dimension condition for the learned representation and its reference measure. Changing the sphere radius therefore does not directly test the curvature dependence in the theorem.

  5. The experimental metric does not establish improved target-domain performance.

    The reported gap is target test error minus source training error. This gap can shrink because training error increases, even when target performance does not improve. The authors acknowledge that smaller radii reduce representation capacity and make training harder, so this is a concrete concern. The experiments do not adequately separate these effects. In addition, the six-domain setting includes MixStyle-generated domains, which are dependent on the original source domains and do not represent additional independent domain draws assumed by the theory.

Additional Concerns

Several references contain incorrect titles, author information, or page ranges. There are also inconsistencies in the proof notation and constants, including Eqs. (33), (39), and (49).

Questions for authors
  1. How is Eq. (48) justified under the definition in Eq. (19)?
  2. Does reducing the radius improve target-domain performance, or does the reported gap decrease because source training error increases?
  3. What establishes the required curvature-dimension condition for the experimental representation?
3 comments

Weaknesses 1: We thank the reviewer for pointing out this error. We confirm that the original manuscript misstated Gouk et al.'s Theorem 2. According to the original paper, one should use its definition of SniS_n^i, namely the set formed by taking one corresponding sample from each source domain, and the domain-level Rademacher complexity over this set. We will correct the notation and text of Theorem 2 and the explanation below Eq. (11) in the revised version.

Regarding the relation to Gouk et al. Eq. (7), we agree that the two are formally similar in the sense of “performing Rademacher averaging over domains as objects”; the original contribution statement that “treating domains as fundamental objects for defining distribution-level complexity” was indeed too strong, and we will weaken it. Our distinction is not that we were the first to define a domain-level Rademacher complexity, but that we explicitly restrict the domain space to a restricted subset of the Wasserstein space, thereby establishing a connection between the geometric properties of this space and domain generalization. Nevertheless, we would still like to emphasize our contribution: this paper is the first to explicitly define the domain space as a Wasserstein space and to introduce Wasserstein geometry into Rademacher-complexity-based domain generalization theory in order to analyze the impact of geometric properties on domain generalization. We believe this is key to revealing what distinguishes generalization across distributions from ordinary generalization, and therefore it remains a unique and prominent contribution to domain generalization.

In summary, we have revised the abstract and the contributions of the original manuscript, as detailed in the revised version of the paper.

Weaknesses 2 and Question 1: Thank you for pointing out our error. We carefully checked the derivation and found that this was caused by the incorrect use of absolute values when defining the domain generalization error in Eqn. (10) and the multivariate risk functional in Eqn. (20). We removed the absolute values from both definitions, which resolves this error. See the revised version of the paper for details (Eqn. (10), Eqn. (20), Eqn. (49)).

Weaknesses 3: We thank the reviewer for this insightful comment. We agree that the curvature-dependent term indeed couples with the Lipschitz constant qzq_z and the domain regularity constant CC, and these quantities need not remain fixed when the representation changes.

To address the coupling with qzq_z, we will remove the Lipschitz assumption on the loss and replace the Kantorovich--Rubinstein duality step in the proof of Theorem 3 with the domain adaptation inequality in Lemma 1 of [1] (Redko et al. (2019)), inheriting its assumptions. This modification eliminates the explicit qzq_z factor from the bound. We will add an appendix summarizing the relevant lemma (Appendix D in the revised paper) and give the revised derivation of Theorem 3 (Eqn. (31) in the revised paper).

To address the coupling with CC, we will no longer treat CC as a fixed constant but keep it as an explicit, representation-dependent factor in the bound, around which we build an explicit invariance interpretation. See the revised version of the paper for details (Definition 4, Eqn. (21), (23), (24)). Guided by this invariance interpretation, we revise the claim about curvature. The original statement that "higher curvature leads to better domain generalization'' is replaced with "domain generalization is jointly affected by the curvature of the representation space and its degree of cross-distribution invariance.'' This statement directly connects our bound to invariant representation learning for domain generalization.

Weaknesses 4 and Question 3: We thank the reviewer for this careful comment. We would like to clarify two points.

  1. Our vMF sampling produces zunitz_{\text{unit}} on the unit hypersphere, so each R⋅zunitR \cdot z_{\text{unit}} lies on the sphere of radius RR centered at the origin, and the soft assignment adds a convex combination of fixed centers cic_i. Hence every feature lies on the union of spheres of radius RR centered at cic_i.

  2. In our setting, the representation space is a Riemannian manifold, the base manifold is a union of hyperspheres equipped with the round metric. For a standard sphere of radius RR, the Ricci tensor is n−1R2g\frac{n-1}{R^2}g where nn is its dimension, RR is the radius, and gg is the Riemannian metric tensor. In our experiments, we set KK to the sectional curvature 1/R21/R^2. Then, by the equivalence theorem given in Theorem 7.3(b) ((1) and (3)) of [2], one can directly verify that our experimental setting satisfies the CD(K,∞)\mathrm{CD}(K,\infty) condition.

Weaknesses 5 and Question 2: We thank the reviewer for these constructive comments. We respond to the two concerns separately.

  1. We thank the reviewer for these detailed concerns. We agree that using a single hypersphere may lead to underfitting, and that the mixture-of-hyperspheres construction was introduced to mitigate this issue but may not eliminate it entirely; indeed, variations in training accuracy are observed in practice. Therefore, we can only verify the relationship between curvature and generalization error by showing how this empirical generalization gap varies, which does introduce a limitation in experimental rigor. We have added explicit discussion of this experimental limitation in the revised version (Appendix G.4).

  2. We agree that MixStyle-generated domains depend on the original source domains and do not correspond to additional independent draws from the meta-distribution assumed by the theory. However, this is a limitation caused by the insufficient number of domains in the datasets, and our experiments with one, two, and three source domains still allow us to verify the effect of the number of source domains on domain generalization.

Additional Concerns: We thank the reviewer for pointing this out, and we apologize for these errors. We will carefully check all references for incorrect titles, author information, and page ranges, and correct them in the revised version. We will also unify the notation and constants throughout the proof, with particular attention to Eqs. (33), (39), and (49), and provide the corrected derivations and constants in the revision.

[1] Redko, Ievgen and Habrard, Amaury and Sebban, Marc. Theoretical analysis of domain adaptation with optimal transport. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 737--753 (2017).

[2] John Lott and C´edric Villani. Ricci curvature for metric-measure spaces via optimal transport. Annals of Mathematics, pp. 903–991, (2009).

Thank you for the detaild response. I appreciate the conribution that: separating domian sampling complexity from the usual within-domain term to add a distinct perspective to existing DG bounds. Although this adds diversity to existing DG bounds, the authors do not provide sufficient evidence to show that the bound is tighter or more explanatory. I decide to keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.