acceptodds
Under review as a conference paper at ICLR 2027

Attribution Ambiguity Can Make Target Bayes Rules Non-Identifiable in EEG Transfer

Abstract

Reliable unsupervised domain adaptation (UDA) for electroencephalography (EEG) decoding depends on information constraining how source and target domains may differ. We ask how far such knowledge can take us: if the laws governing how latent neural components transform across domains were fully specified, would that be enough to determine the optimal target classifier? We examine this question in an idealized Oracle setting in which the full source distribution of observations and labels, the target observation marginal, and the corresponding component-specific operators are known exactly, while other common sources of uncertainty are removed. Yet a simple Gaussian construction yields a negative answer. We trace this failure to attribution ambiguity: scalp EEG observations reflect aggregate neural activity but not how predictive activity is partitioned across component types. Although each type's cross-domain operator is known, allocations indistinguishable in the source domain can remain consistent with different target observation–label relationships that require conflicting optimal decisions. To characterize the resulting learning barrier, we derive an information-theoretic lower bound showing that it cannot be overcome by algorithmic choice alone. With additional structure, a complementary upper bound shows that the target decision can be learned without full latent-attribution recovery. Finally, EEG simulations instantiate the theoretical mechanism, while real-EEG experiments provide complementary evidence for its learner-side predictive implications.

Primary area:
learning theory
Keywords:
domain adaptation, transfer learning, identifiability, learning theory, distribution shift, electroencephalography (EEG) transfer learning

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

This paper investigates why unsupervised transfer learning for EEG may fail. The authors propose that, even with complete knowledge of how each component changes from the source domain to the target domain, it may still be impossible to identify the optimal classifier for the target domain. The paper first demonstrates this phenomenon through a mathematical counterexample and then derives a lower bound on the classification risk that no algorithm can avoid. The authors further prove that certain learning guarantees can be established when additional information, such as covariance predictions and classification directions, is available. Finally, the authors investigate this phenomenon through experiments on both simulated and real EEG data.

Critical strengths

The paper is generally well-written, the figures and tables are clearly presented, the research question is well-defined, and the examples are easy to understand.

Critical weaknesses

My main concern is that the connection between the EEG research motivation and the core theoretical contributions is not sufficiently strong:

The authors limit their study to the EEG domain, but I think the problem they investigate is actually a general one. The paper simply uses EEG as the application scenario throughout, rather than formulating the research question to specifically address a particular challenge in this field. Moreover, none of the mathematical conditions in the formulas seem to be unique to EEG. Therefore, it seems that the paper could have discussed a general problem. If so, then, in my view, many of the authors' claimed contributions cannot be clearly distinguished from existing work. See the questions below for details.

Questions for authors
  1. Is Theorem 1 merely a direct application of the classical Le Cam two-point lower bound (Tsybakov, 2009), with the loss separation expressed as (C_{\mathrm{conf}})? Given that its proof does not use the EEG-specific assumptions or the additive mixture structure, what independent theoretical contribution does this theorem provide?

  2. The conditions of Theorem 2 do not seem to depend on EEG-specific properties, and related methods have already been studied in previous work (Minimally Informed Linear Discriminant Analysis: Training an LDA Model with Unlabelled Data, Heintz et al., 2023). Could the authors answer the following question: What new guarantees does Theorem 2 provide compared with existing theories on unlabeled LDA and Gaussian classification, such as Heintz (2023)?

  3. Is Proposition 1 essentially a reparameterization of existing domain adaptation non-identifiability results? In its scalar construction, XS=J1+J2X_S=J_1+J_2 and XT=2J1+J2X_T=2J_1+J_2 imply J1=XT−XSJ_1=X_T-X_S and J2=2XS−XTJ_2=2X_S-X_T. Thus, without additional restrictions on the latent components, arbitrary source–target couplings can be represented using the same fixed operators. Does the known-operator assumption impose any substantive constraint that prevents the classical counterexamples of Ben-David et al. (2010) from being embedded in this model?

  4. What evidence shows that the non-identifiability described in Proposition 1 occurs in real EEG transfer tasks?

8 comments
AuthorsOpenReview ↗

Official Response to Reviewer moos [1/5]

We appreciate your thorough review and positive feedback on the clarity of our research question, examples, figures, and tables. Your penetrating questions on the EEG motivation, theoretical novelty, and empirical scope have helped us sharpen the manuscript.

Overarching Concern: Connection between EEG motivation and general theory. Our study builds a two-way link between the physical structure of EEG generation and general learning theory:

  1. From the perspective of EEG. EEG does not merely motivate our study; it provides the scientific origin of our research question, its underlying generative structure, and the scope of our results. Existing EEG transfer research has extensively investigated inter-subject differences and developed methods to compensate for them. Our work instead asks a more fundamental question: why can EEG transfer remain difficult in the first place? We seek to identify and characterize a concrete generative mechanism that can give rise to this difficulty. Without the scientific question and additive-aggregation observation structure provided by EEG, our theoretical investigation would lose both its physical grounding and the structural basis for its applicability.
  2. From the perspective of UDA. Learning theory provides the tools to characterize the learning consequences of changes in EEG's generative mechanisms and to generalize the resulting insights. Existing UDA theory has established the importance of assumptions and knowledge about cross-domain changes for reliable transfer. However, general characterizations of distribution shifts and their learning consequences do not, by themselves, explain the concrete generative mechanisms from which those shifts arise. Starting from EEG's generative structure, we use learning theory to characterize the consequences of a specific transfer-failure mechanism and extend the resulting insights to other systems sharing this structure. Without this theoretical framework, our investigation of EEG's generative mechanisms would lack a rigorous characterization of their decision consequences and general learning limits.
  3. Significance and novelty. To the best of our knowledge, this is the first learning-theoretic study to identify a concrete mechanism of transfer failure rooted in the physical structure of EEG generation and to derive decision-theoretic bounds on the resulting excess risk. This analysis thus bridges empirical research on performance degradation in EEG transfer and theoretical work on the limits of UDA.

Below, we address your four questions to explain the mechanism, clarify what our theoretical results add to prior work, and show what the EEG evidence supports.

AuthorsOpenReview ↗

Official Response to Reviewer moos [2/5]

Is Theorem 1 merely a direct application of the classical Le Cam two-point lower bound (Tsybakov, 2009), with the loss separation expressed as CconfC_{\mathrm{conf}}? Given that its proof does not use the EEG-specific assumptions or the additive mixture structure, what independent theoretical contribution does this theorem provide?

We appreciate the reviewer's thoughtful question, which gives us a welcome opportunity to clarify both the role of the Le Cam technique in Theorem 1 and its connection to EEG's generative structure.

  1. The generality of the proof does not imply a disconnect from the EEG mechanism. We distinguish two claims: that Theorem 1's lower-bound proof does not explicitly use EEG-specific assumptions or the additive-mixture structure, and that Theorem 1 is unrelated to these structures. The former is correct; the latter does not follow. EEG provides the scientific origin of our research question, the underlying additive-aggregation structure, and the structural scope of the mechanism we study. The two conflicting worlds used to instantiate the lower bound are not postulated independently of this structure; they arise from attribution ambiguity within the EEG-motivated generative model (Proposition 1). Once this mechanism has been shown to produce incompatible target-optimal decisions, Theorem 1 characterizes its learning consequences at a general decision-theoretic level. The absence of EEG-specific assumptions from the proof therefore reflects the abstraction of a structurally grounded learning obstruction, not a disconnect from its scientific origin.
  2. Determining the right decision-theoretic separation functional is non-trivial. We agree that the coupling step in Theorem 1 uses the standard two-point total-variation argument (Tsybakov, 2009). However, applying it here first requires establishing which differences between Oracle-compatible models can actually force excess risk. Appendix A.7.1 (Proposition A.1) rigorously proves that neither structural non-identifiability necessarily implies risk non-identifiability, nor risk non-identifiability necessarily implies decision conflict. These strict separations are established within the same additive generative family, under identical Oracle information and fixed component-specific transformations: different latent attributions can yield identical risk functions, while different risk functions can still share the same Bayes-optimal decision. Definition 1 and Lemma 1 then characterize the decision conflict genuinely relevant to learning, enabling Theorem 1 to quantify its unavoidable cost under protocol-dependent distinguishability. The classical two-point argument provides the lower-bound technique; identifying the decision-relevant consequences of attribution ambiguity is a distinct theoretical task.
  3. Classical proof techniques do not diminish the novelty of a newly characterized learning barrier. The use of a classical proof technique does not, by itself, determine whether the learning phenomenon being characterized or the resulting theoretical understanding is novel. Learning theory advances not only through new mathematical tools, but also by turning domain-specific learning phenomena that lack rigorous explanations into precise scientific problems whose mechanisms and performance limits can be characterized. Our work follows this direction by uncovering an information-theoretic source of learning difficulty in EEG transfer.
AuthorsOpenReview ↗

Official Response to Reviewer moos [3/5]

Is Proposition 1 essentially a reparameterization of existing domain adaptation non-identifiability results? In its scalar construction, XS=J1+J2X_S = J_1 + J_2 and XT=2J1+J2X_T = 2J_1 + J_2 imply J1=XT−XSJ_1 = X_T - X_S and J2=2XS−XTJ_2 = 2X_S - X_T. Thus, without additional restrictions on the latent components, arbitrary source–target couplings can be represented using the same fixed operators. Does the known-operator assumption impose any substantive constraint that prevents the classical counterexamples of Ben-David et al. (2010) from being embedded in this model?

Thank you for pointing out the coupling freedom in the scalar construction.

  1. The algebraic observation is correct. We agree that known operators alone do not exclude arbitrary source–target couplings in the unrestricted scalar construction, and therefore do not rule out classical UDA counterexamples. 🔵 We have clarified in Section 3.2 that Proposition 1 establishes the existence of incompatible target Bayes decisions under identical Oracle information, with the scalar construction serving as a transparent minimal witness of attribution ambiguity.
  2. The decision conflict persists under substantive generative restrictions. The conflict exhibited by Proposition 1 does not rely on arbitrary coupling freedom. 🔵 To make this explicit, we have added a constrained EEG realization of Proposition 1 in Appendix A.2 and summarized it in Section 3.2. This construction fixes the EEG lead field, component-specific rotations, and residual covariances, while requiring conditionally independent, nondegenerate Gaussian components. The two worlds differ only through a one-dimensional reallocation of predictive activity, while their joint source–target residual law is fixed rather than arbitrary. Nevertheless, they share identical Oracle information, and their full-observation Bayes rules are each unique almost everywhere and opposite almost everywhere.

Therefore, the general impossibility results of Ben-David et al. (2010) do not by themselves establish decision conflict under our stronger Oracle information and restricted generative assumptions. Their specific disjoint-support constructions (Theorems 1–2) also cannot be directly embedded in this constrained Gaussian subclass.


Significance: The Oracle Information Limit and a Necessary Test for Target-Decision Recovery

The constrained construction has two further implications.

  1. The information limit of the Oracle protocol. The Oracle gives the learner the complete source observation–label distribution and target observation marginal, together with exact component transformations. Yet even this information is insufficient to determine the target Bayes rule. Under an equal prior over the two candidate worlds, one bit of uncertainty remains about which world is true. Further labeled-source or unlabeled-target samples under the same protocol cannot resolve it.

  2. A necessary test for exact target-decision recovery. Existing theories give conditions under which recovery is possible. Attribution ambiguity raises a question that any such guarantee in our setting must address. Do its additional information or structural assumptions resolve the conflict between target Bayes decisions? Exact recovery across the remaining compatible models requires that they admit a common Bayes-optimal rule.

Thus, our contribution goes beyond demonstrating non-identifiability: it identifies a concrete decision-sufficiency requirement that positive Bayes-recovery guarantees must satisfy in additive-aggregation systems, while deepening our understanding of the generative mechanisms underlying transfer difficulties in these systems.

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

The paper addresses the problem of unsupervised domain adaptation. The model and setting the authors study is motivated by the EEG literature and is a relatively simple model.

The authors claim that the main contribution of the paper is the description of the non-identifiability of the Bayes rule even when all information about a transformation is known.

The authors show that the ambiguity can be lessened in the case where the problem contains more structure.

The authors do relatively light empirical investigations.

Critical strengths

Unsupervised domain adaptation is a very important problem. Learning a decision function from scratch is often not possible and so leveraging any / all information available is often necessary. While dense, the paper's description of their setting is technically sound.

Critical weaknesses

The writing is very dense and hard to follow. The authors fail to mention that a Bayes optimal decision function is just one element of the argmin of the risk and, as such, is never guaranteed to be unique. Given this, I do not view the author's stated main contribution of describing this ambiguity in a particular context very interesting or novel.

Questions for authors

Apart from the setting's relevance to modeling EEG data, why is this non-identifiability / non-uniqueness of a Bayes' decision function particularly interesting?

3 comments
AuthorsOpenReview ↗

Official Response to Reviewer Fqds [1/1]

The writing is very dense and hard to follow. The authors fail to mention that a Bayes optimal decision function is just one element of the argmin of the risk and, as such, is never guaranteed to be unique. Given this, I do not view the author's stated main contribution of describing this ambiguity in a particular context very interesting or novel.

Thank you for reviewing our work. We first clarify the distinction underlying the concern about Bayes-optimal non-uniqueness.

  1. The two forms of ambiguity are fundamentally different. The reviewer refers to the possibility of multiple Bayes-optimal classifiers within a single model. Our work studies whether different models with identical available information can require incompatible optimal decisions.
  2. Proposition 1 explicitly rules out within-model non-uniqueness as the explanation. Proposition 1 (Section 3.2, p. 4) constructs two models with identical Oracle information whose target Bayes rules are each unique almost everywhere, yet opposite almost everywhere.
  3. Definition 1 explicitly accounts for non-unique minimizers. Definition 1 (Section 4, p. 5) defines pairwise Bayes conflict through the minimum combined Bayes excess risk of using a common classifier across two compatible models. This formulation does not assume that Bayes-optimal classifiers are unique.
  4. Lemma 1 rigorously characterizes when decision conflict occurs. Lemma 1 (Section 4, p. 5) establishes that conflict is positive precisely when the sets of Bayes-optimal actions in the two models have no common action on a set of positive target probability. This explicitly accommodates non-unique optimal actions. Figure 1(b) further illustrates the cross-model conflict.

The possibility of non-unique Bayes minimizers is therefore explicitly addressed in our manuscript and does not explain or invalidate the decision conflict established by Proposition 1.

Novelty and significance. EEG provides the scientific origin and generative structure of our research question, while learning theory enables us to characterize its decision consequences and extend the resulting insights to other systems sharing this structure. To the best of our knowledge, this is the first learning-theoretic study to identify a concrete mechanism of transfer failure rooted in the physical structure of EEG generation and to derive decision-theoretic bounds on the resulting excess risk. 🔵 Our new constrained Gaussian construction further demonstrates that this conflict persists under substantive generative restrictions. Our study as a whole bridges empirical research on performance degradation in EEG transfer and theoretical work on the limits of UDA.

Our responses to the other reviewers discuss related points in more detail. We welcome your feedback and would be happy to answer any further questions.

Reviewer_FqdsOpenReview ↗

reply

Thank you for taking the time to respond.

Beyond the denseness of the presentation, the fundamental issue I have with the paper is that are many ways for a Bayes decision rule to be non-identifiable and I do not think the current paper properly contextualizes the stated contribution with this property.

Further, for a very general purpose community like ICLR I think the contribution of identifying a particular failure mode of a particular model from a particular (relatively niche) subfield is not likely to have a large impact (I could be wrong!). In the context of ICLR, the paper would benefit from finding additional settings where the studied model is relevant / reasonable.

I have kept my original score.

AuthorsOpenReview ↗

Request for Clarification of Reviewer Fqds’s Remaining Theoretical and Significance Concerns

We sincerely appreciate your time and consideration in reviewing our work and providing further feedback. We would welcome further clarification of your central concern, as the objections in your original review and subsequent response appear to address different theoretical issues.

1. Clarification of the theoretical objection

In your original review, you stated that the authors “fail to mention” that a Bayes-optimal decision function is an element of the risk minimizer set and is not guaranteed to be unique. This concerns the possible multiplicity of Bayes-optimal solutions within a single model. Our previous response addressed this point directly.

In your follow-up, however, you identify the fundamental issue as the existence of “many ways for a Bayes decision rule to be non-identifiable.” This is a substantially broader statement. It does not specify whether your concern remains the non-uniqueness of minimizers within a model, or instead concerns the non-identifiability of optimal decisions across models compatible with the learner's information. These are distinct theoretical issues, but your follow-up does not clarify which one remains unresolved or how it relates to your original objection. We would therefore appreciate clarification of whether you regard the original objection as unresolved and, if so, how our previous response fails to address it.

2. Request for specific, manuscript-grounded concerns

We agree that many forms of non-identifiability have been established in the literature. Different studies have made substantive contributions to the field by characterizing the principles, mechanisms, and consequences of non-identifiability in their respective settings. Our work contributes to this line of research by characterizing Bayes-decision non-identifiability arising from attribution ambiguity in additive EEG transfer, together with its decision-theoretic consequences.

To help us address your concern substantively, could you identify a particular prior result that you believe already establishes one of our principal conclusions under comparable assumptions, or specific theoretical arguments, assumptions, or experimental findings in our manuscript that you consider insufficient? We would also appreciate knowing which forms of Bayes-decision non-identifiability you consider most relevant and what relationship or distinction you believe our paper should establish with them. Such clarification would allow us to examine the relevant issues directly, rather than speculate about the comparison underlying your assessment.

3. Clarification regarding ICLR scope and significance

We also appreciate your concern about the potential breadth of the contribution. The official ICLR 2027 Call for Papers explicitly includes both “applications to neuroscience & cognitive science” and “learning theory” among its subject areas. Our study lies at the intersection of these two areas. It originates from a problem grounded in the physical structure of EEG generation and develops its implications through general learning-theoretic analysis. To the best of our knowledge, this is the first learning-theoretic study to identify a concrete mechanism of transfer failure rooted in the physical structure of EEG generation and to derive decision-theoretic bounds on the resulting excess risk.

We recognize that being within the conference's stated scope does not, by itself, establish sufficient significance. Given the scientific and theoretical contributions described above, could you clarify whether your primary concern is their significance for EEG transfer itself, or whether their relevance to the broader ICLR community has not been sufficiently established? We would welcome your perspective on which aspects require stronger justification. We also agree that demonstrating relevance in additional settings could further broaden the impact. Related additive observation structures arise in other domains, including MEG and acoustic source mixtures. Your suggestion has prompted us to consider whether the Discussion could more explicitly address these potential extensions, including their underlying assumptions and limitations.

We have provided substantive mathematical clarification in response to the original review. We would greatly appreciate further clarification of the remaining objections, so that we can engage with the precise theoretical or scientific concerns underlying your assessment.

Thank you for your time and consideration.

Reviewer_v2JJOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether exact knowledge of cross-domain transformation structure is sufficient to determine the target-optimal prediction rule in unsupervised domain adaptation (UDA) for EEG decoding. The authors consider an idealized Oracle setting in which the full source observation-label distribution, target observation marginal, and component-specific transformation operators are known exactly. They identify a phenomenon termed attribution ambiguity, where scalp EEG observations reflect aggregate neural activity but do not reveal how predictive activity is partitioned across latent component types. Through a Gaussian construction, the authors show that two generative models compatible with the same Oracle information can have opposite target Bayes rules. To characterize this learning barrier, they derive an information-theoretic lower bound on unavoidable Bayes excess risk and quantify how labeled target samples progressively resolve the ambiguity. Under additional structure, a complementary finite-sample upper bound shows that the target decision can be learned without full latent-attribution recovery. Finally, controlled EEG simulations instantiate the theoretical mechanism under forward-model constraints, while real-EEG experiments provide complementary evidence for its learner-side predictive implications.

Critical strengths
  1. The paper investigates a theoretical problem by examining target Bayes-rule non-identifiability even when component-specific cross-domain transformations are fully known.

  2. The theoretical analysis is comprehensive, combining a Gaussian counterexample, an information-theoretic lower bound, and a complementary finite-sample recovery guarantee.

  3. The paper provides extensive experiments, including controlled EEG simulations and real-EEG analyses, with additional controls supporting the proposed attribution ambiguity mechanism.

Critical weaknesses
  1. The incremental novelty and significance relative to existing UDA identifiability and impossibility results could be more clearly established. While the known component-specific transformations distinguish the setting from prior work, the extent of the additional theoretical contribution beyond existing non-identifiability principles and standard information-theoretic arguments remains unclear.

  2. The finite-sample guarantee relies on restrictive assumptions, particularly accurate moment predictions and a positive reference-alignment condition. The latter excludes the decision-conflicting sign-reversed example constructed earlier and cannot be verified from unlabeled target data alone, limiting the applicability of the guarantee.

  3. The empirical evidence remains indirect relative to the theoretical claims. The simulations evaluate classifier exchange loss rather than Bayes conflict, while the real-EEG experiments establish associations between learner-side attribution differences and target prediction divergence. The proposed finite-sample recovery guarantee is also not evaluated experimentally.

Questions for authors

See weaknesses. Given the extensive literature on UDA non-identifiability and impossibility results, could the authors clarify to what extent existing identifiability principles and standard information-theoretic techniques can be applied to the proposed additive latent structure? In particular, what additional theoretical insights do the results in this paper provide beyond these existing approaches?

6 comments
AuthorsOpenReview ↗

Official Response to Reviewer v2JJ

We sincerely thank the reviewer for the positive and constructive assessment of our work. We particularly appreciate that you evaluated our interdisciplinary study as a whole, recognizing how its EEG motivation, central information-sufficiency question, theoretical analysis, and empirical investigation fit together. We value your thoughtful questions on theoretical novelty, recovery assumptions, and empirical support, and address each individually below.

AuthorsOpenReview ↗

Official Response to Reviewer v2JJ [1/3]

See weaknesses. Given the extensive literature on UDA non-identifiability and impossibility results, could the authors clarify to what extent existing identifiability principles and standard information-theoretic techniques can be applied to the proposed additive latent structure? In particular, what additional theoretical insights do the results in this paper provide beyond these existing approaches?

The incremental novelty and significance relative to existing UDA identifiability and impossibility results could be more clearly established. While the known component-specific transformations distinguish the setting from prior work, the extent of the additional theoretical contribution beyond existing non-identifiability principles and standard information-theoretic arguments remains unclear.

We address Weakness 1 and Question 1 jointly below, as both concern how existing UDA identifiability principles and information-theoretic tools apply to our setting and what additional insights our results provide.

1. Novelty and Significance.

Within ICLR's scope, learning theory and applications to neuroscience & cognitive science are two distinct subject areas, and our work sits at their intersection. While EEG transfer remains heavily reliant on target calibration, existing UDA theory does not explain how EEG's physical generative structure creates transfer barriers. To the best of our knowledge, this is the first learning-theoretic study to identify a concrete transfer-failure mechanism rooted in EEG generation and derive decision-theoretic bounds on its unavoidable excess risk, bridging empirical EEG transfer research and UDA theory.

Manuscript revisions: 🔵The revised Introduction and Section 7 Discussion and Conclusion clarify the contribution and its implications for target-decision recovery. 🔵Appendix D.6 discusses how the mechanism may arise in other additive recording systems.

2. Additional Theoretical Insights.

We expose a two-stage barrier that any theory of reliable transfer in additive systems must overcome:

  • (1) Information Barrier: Exact cross-domain laws still fail to identify the target rule. Prior UDA impossibility results rely on unknown cross-domain shifts. We push the problem to its information extreme instead. Specifically, we grant the learner exact component-specific transformations, along with complete source joint and target marginal laws (Section 3). Yet the target decision still lacks one bit of information (Proposition 1; Appendix A.2). Two scenarios can look identical under Oracle knowledge, yet require opposite label predictions. Consequently, every algorithm faces a strictly positive excess-risk lower bound (Theorem 1; Appendix A.4). Attribution ambiguity thus sets a necessary test for positive recovery theories: how do their extra assumptions resolve the decision conflict that survives Oracle knowledge?
  • (2) Statistical Barrier: Population identifiability does not guarantee sample feasibility. Extra structure can pass this test and make the target rule identifiable in principle (Theorem 2). Even then, observable evidence separating competing rules can be severely weakened. 🔵 Our new minimax analysis (Proposition 3 in Section 5; Theorem A.2 in Appendix A.8.6) pinpoints background masking as a concrete cause. Background masking cancels discriminative cues in the unlabeled target covariance without changing the target Bayes risk. As a result, optimal unlabeled-target sample complexity degrades from Θ(r−2)\Theta(r^{-2}) to Θ(r−4)\Theta(r^{-4}). In plain terms, a rule that is learnable on paper can still require far more unlabeled target data than an EEG experiment can collect.

3. Applicability of Existing Positive Identification Theories.

Existing positive identification frameworks cannot resolve the Bayes-decision conflict under our Oracle information alone because their core premises fail in single-source additive EEG transfer: scalp mixing (X=GJX = GJ) is many-to-one rather than invertible (ruling out Kong et al., 2022), only one labeled source domain is available (ruling out multi-source identification in Ng et al., 2025), competing marginal-matching maps can imply opposite decisions (Gulrajani & Hashimoto, 2022), and unlabeled target covariance recovery still leaves class sign orientation undetermined without an aligned source reference (Heintz et al., 2026). 🔵 Our revised Appendix D (Sections D.2–D.4), currently being finalized for our upcoming revised PDF, details these boundaries.

References

  • Gulrajani & Hashimoto (2022). Identifiability Conditions for Domain Adaptation. ICML.
  • Heintz, Francart, & Bertrand (2026). Minimally Informed Linear Discriminant Analysis: Training an LDA Model with Unlabelled Data. Signal Processing, 239, 110226.
  • Kong et al. (2022). Partial Disentanglement for Domain Adaptation. ICML.
  • Ng et al. (2025). A General Representation-Based Approach to Multi-Source Domain Adaptation. ICML.
AuthorsOpenReview ↗

Official Response to Reviewer v2JJ [2/3]

The finite-sample guarantee relies on restrictive assumptions, particularly accurate moment predictions and a positive reference-alignment condition. The latter excludes the decision-conflicting sign-reversed example constructed earlier and cannot be verified from unlabeled target data alone, limiting the applicability of the guarantee.

The proposed finite-sample recovery guarantee is also not evaluated experimentally.

We thank the reviewer for raising two closely related concerns about Section 5: the restrictive assumptions underlying Theorem 2 and the absence of its experimental evaluation. Both questions concern the role and applicability of our positive recovery theory, which we clarify together below.

1. The information conditions for decision recovery.

We agree that these recovery conditions are restrictive. However, Section 5 is a theoretical and information-theoretic inquiry into what conditions could make decision recovery possible. It is not intended as a direct method for real-world data recovery. Because Oracle information alone cannot rule out opposite target predictions (Sections 3 and 4), any recovery guarantee must specify extra structure or side information to resolve this conflict. Our analysis makes one set of sufficient conditions explicit, while practical recovery in real EEG remains an open problem. 🔵 To make this theoretical role clear, we shortened the description of the learner in the main text and retained the full learning rule in Appendix A.6.2.

2. The role of Theorem 2 and the scope of its guarantee.

Within this framework, Proposition A.2 establishes population decision recovery, and Theorem 2 extends the result to finite samples with an explicit Bayes excess-risk guarantee. We acknowledge that Theorem 2 has not been experimentally evaluated and that establishing these conditions in real EEG remains a separate challenge.

3. From decision recovery to statistical feasibility.

🔵 To clarify the broader purpose of Section 5, we have reorganized it as Conditions and Sample Requirements for Target Decision Recovery and added Proposition 3 and Theorem A.2 (Appendix A.8.6). These results address a further question: even when extra structure makes a target decision identifiable in principle, can it be learned with a feasible amount of data?

🔵 Our new minimax analysis identifies background masking as a concrete mechanism that governs this statistical limit. When background covariance shifts cancel discriminative cues in the unlabeled target covariance, the observable evidence weakens without changing the target Bayes risk. In our Gaussian benchmark, at 75% Bayes balanced accuracy and a one-percentage-point tolerance in expected excess risk, the minimum requirement rises from 71 unlabeled target trials without masking to 2,634 under complete masking. Thus, statistical costs can prevent practical recovery even when the target rule is identifiable at the population level.

Section 5 thus identifies the additional information needed for target-decision recovery, establishes finite-sample guarantees under explicit conditions, and quantifies how statistical costs shape recovery under limited data budgets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.