acceptodds
Under review as a conference paper at ICLR 2027

Calibration or Location? Predictive Uncertainty under Gauge Confounding

Abstract

Representation learning often aims to remove sensor or acquisition variation while preserving task information. These goals conflict when nuisance and task transformations induce the same observation: nuisance invariance then collapses the corresponding target orbit. For coherent arrays, we prove that an unknown sensor phase affine in position is observationally equivalent to a change in incident wavevector. Neither raw I/Q nor conjugate-phase features can then identify absolute direction from one snapshot. Lower bounds for point risk and prediction-set width require valid predictions to represent this ambiguity unless additional information breaks the gauge. We test these implications through controlled interventions on public 77 GHz raw ADC data. A prespecified label-only pipeline retains 281 automotive frames. Across 843 injected phase ramps, every retained beam peak translates as predicted: mean circular spatial-frequency error rises from 0.042 to 0.144 and returns to 0.042 with oracle correction. An independent audit confirms complete-spectrum translation and noncoherent range-Doppler invariance. In learned benchmarks, three five-model ensembles achieve 88.6–90.2% coverage at the training drift bound but only 6.1–7.1% when the test bound is sixteen times larger, while widths remain nearly constant. Under gauge-paired worlds, byte-identical inputs leave every prediction unchanged and shifted coverage falls to zero. Given a specified nuisance law, an analytic posterior and a learned orbit head recover about 90% coverage by widening along the unidentified orbit; a noisy known-direction anchor restores 90.4%. The theorem covers arbitrary geometries under narrowband far-field propagation; the learned study uses a single-source ULA without multipath.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
uncertainty quantification, predictive uncertainty, identifiability, gauge symmetry, invariant representation learning, distribution shift, sensor calibration, coherent arrays, radar localization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_2mkfOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As it is hard to read the paper it maybe reflected in the following summary. The goal is learning representation in presenece of nuisance parameter that affect the task performance. The paper starts with general problem definition and several theoretical results in general setting. The paper demonstrates the of array ambiguity where the direction of arrival (the label) is interpreted by phase (nuisance parameter). Then the paper discusses the prediction of uncertainty with confidence interval.

Critical strengths

As it is hard to read the paper maybe there are additional strengths. This work address an interesting example of array ambiguity and representation learning, can provide an interesting challenge to machine learning community.

Critical weaknesses

(1) Novelty: Array ambiguity has a long line of works in the signal processing community [R.1, R.2, R.3] and it unclear what is the contribution of Theorem 3 and ULA corollary as they are study widely. Moreover, it is unclear how the authors suggest to address this and what is the underline method and relation to representation learning. In addition Theorem 2 is trivial and a direct results of the assumptions. Finally, it is unclear from the text what are the main contributions.

(2) Clarity of presentation: it is very hard to read the paper and try to have a flow of understanding of the paper. Here are few critical points why this paper is unclear:

1.1 Figures are not referenced in the text and the reader needs to relate the figure to the part of the text is relevant .

1.2 There is a lack of related work or list of contributions that would specify what is the contribution of this works.

1.3 There is a mix of notation in the text and it unclear how to relate example to the notation at the beginning for example the y label and h nuisance parameter. (e.g. equation 3-5 what is nuisance and what is label? )

1.4 There is no clear statement of what are the contributions of this work.

1.5 There is not reference from the main text to proofs, and the reader need to search for them in the appendix.

1.6 Figure legends are not specified or unclear (e.g. Figure 4 black dashed line)

The notations are hard to follow, for example:

1.1 In equation 3 and 4 is x is the same as x_{a}

1.2 In equation 3, A is not defined (I assumed it is the number of sensors)

1.3 Mismatch in indexing between equations (3) and (5)

1.4 Several acronyms are not defined such as CQR, ADC, ULA, etc.

(3) Experiment Results: There is no baseline compression and it hard to understand if there is any new observation from the results the enable better understanding of the problem.

[R.1] A. J. Weiss and B. Friedlander, "Eigenstructure methods for direction finding with sensor gain and phase uncertainties," Circuits Syst. Signal Process., vol. 9, no. 3, pp. 271–300, 1990

[R.2] E. K. L. Hung, "A critical study of a self-calibrating direction-finding method for arrays," IEEE Trans. Signal Process., vol. 42, no. 2, pp. 471–474, Feb. 1994.

[R.3] Liu, Zhexian, et al. "Ambiguity-Free DOA Estimation for Linear Arrays via Sparse Bayesian Phase Error Self-Calibration." 2025 9th International Conference on Vision, Image and Signal Processing (ICVISP). IEEE, 2025.

Questions for authors

(1) What the authors see as the main contributions of this work?

(2) As array ambiguity has a long line of works in signal processing community, can the authors state what is the novelty or new in Theorem 3 and ULA corollary?

(3) How are the different experiment related ?

(4) In Array model equations 3-5 what are the labels Y usually is the direction of arrival but it unclear from the text?

(5) What is the spatial dimension? p=3? or p=2?

(6) In line "Our posterior, oracle, and anchor are matched nonlearned references;" where is this compression shown?

2 comments

Thank you for the detailed list. We are sorry the submitted version was hard to read. Most of your points are addressed in the revision.

Q1. Main contributions. The introduction now lists three: (i) known limits, restated as a precision requirement for a requested interval width; (ii) a map from each kind of calibration information to the coverage claim it supports, with learned evidence; (iii) a test of reference transfer, in simulation and on real recordings.

Q2. Novelty of the gauge theorem and the ULA corollary. None. They are the classical ambiguity of array self-calibration, written for general geometry. The revision labels the theorem "classical", cites Weiss and Friedlander (1990), Wylie et al. (1994) and Hung (1994) next to it, and makes no novelty claim. The result you describe as a direct consequence of the assumptions is now stated as an elementary proposition.

Q3. How the experiments relate. Section 3 follows the rows of Table 1 in order: input alone, a law, average versus fixed-drift coverage, a bound, a labelled panel. Section 4 treats the last row, a shared reference.

Q4. Labels and nuisance in the array model. The text after Eq. (3) now states that the label is the wavevector coordinate k (the direction, for a line), the nuisance is the sensor phase calibration, and the forward operator maps both to the observation.

Q5. Spatial dimension. p = 1 for a line and p = 2 for a plane. All learned experiments use p = 1. This is stated where the array model is introduced.

Q6. The sentence on matched non-learned references. That sentence has been removed. The comparison it referred to is Table 12 (Appendix C2), which lists every method with the information it uses, its coverage and its width.

Notation and presentation.

  • A is defined as the number of sensors, and x is defined as the vector with components x_a, indexed from 0 to A−1 in both Eq. (3) and the ULA corollary.
  • ULA, ADC, CQR and the other acronyms are defined at first use. Appendix A1 has a notation table.
  • Each theorem and proposition is followed by a pointer to its proof in Appendix A2.
  • Every figure and table in the main text is referenced where it is discussed, and Section 5 covers related work.

Baselines. Narrow training, wide augmentation, a supplied law, a conservative set, a labelled panel with direct and orbit heads, and an oracle are compared in Tables 2 and 12 (an excerpt of Table 12 is in our general comment above); label-only conformal calibration with the same labels as the anchors is compared in Appendix B3. We agree the submitted version did not make this easy to find.

How we address the ambiguity. We do not propose a representation-learning method that removes it; the results show that none can from the input alone. The paper states which additional information supports which coverage claim.

Thanks the authors for the detailed response. However, following the authors acknowledgment three is limited theoretical novelty (gauge theorem) and the other reviewers comments, I will keep the current score.

Reviewer_p6M2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the ambiguity between sensor calibration errors and target localization in coherent arrays. It shows that certain phase calibration errors are indistinguishable from changes in target direction, making accurate localization and uncertainty estimation impossible from a single observation. The authors derive theoretical bounds on prediction error and uncertainty, and propose accounting for calibration ambiguity through nuisance-aware prediction intervals or additional calibration information. Experiments on synthetic data and controlled phase perturbations of real radar ADC measurements validate the ambiguity and show that standard uncertainty estimation methods become overconfident under calibration shifts, while the proposed alternatives improve coverage.

Critical strengths
  • The problem is clearly formulated. The paper gives a clean description of the ambiguity between calibration phase and target direction, and the theoretical setup is easy to follow.

  • The experiments are consistent with the theoretical construction. The controlled phase-ramp tests reproduce the predicted behavior, although the experimental scope is quite limited.

Critical weaknesses
  • The core phase-calibration/DOA ambiguity is already well known in array self-calibration. The paper mainly moves this known issue into an uncertainty-estimation setting, so the conceptual advance feels incremental.
  • The exact ambiguity requires affine phase drift under a narrowband far-field model; the learned experiments further simplify this to a single-source ULA. It is unclear how representative this is of real radar calibration errors.
  • The authors inject artificial phase ramps into real ADC data and then verify the expected shift. This shows the algebra is correct, but does not demonstrate that the same issue actually occurs in deployed radar systems.
  • Most results use synthetic single-source data and small MLPs, without realistic multi-target scenes, multipath, temporal context, or modern radar perception models[3]. This makes the broader claims about learned radar uncertainty hard to support.
  • Recovering coverage requires the test-time nuisance law, calibration bound, or a known-direction anchor. In practice, these quantities are often exactly what is unknown, so the method does not solve the more realistic setting.
  1. Wylie, Marilynn P., Sumit Roy, and Hagit Messer. "Joint DOA estimation and phase calibration of linear equispaced (LES) arrays." IEEE Transactions on Signal Processing 42.12 (1994): 3449-3459.
  2. Ramamohan, Krishnaprasad Nambur, et al. "Self Calibration of Scalar and Vector Sensor Arrays." arXiv preprint arXiv:2104.02561 (2021).
  3. Rebut, Julien, et al. "Raw high-definition radar for multi-task learning." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.
  4. Liu, Yang, et al. "Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion." Advances in Neural Information Processing Systems 36 (2023): 53964-53982.
Questions for authors
  • How common is the assumed affine phase drift in real automotive radar systems? Do you have any evidence that this failure mode occurs naturally, rather than only under injected perturbations?
  • Can you clarify how the main conclusions change outside the narrowband far-field setting, especially with near-field propagation, multipath, and more general channel errors?
  • The learned experiments use synthetic single-source data and small MLPs. Do you have evidence that the same uncertainty failure appears in a realistic multi-target radar perception model or on real calibration drift?
  • Your mitigation requires the nuisance law, calibration bound, or a known-direction anchor at test time. What happens when none of these are available or accurately known?
1 comment

Thank you for the precise questions. We answer them directly.

Q1. How common is affine drift in real automotive radar? Is there evidence it occurs naturally? We have no such evidence, and the revision says so. Public measurements report scalar phase changes; if their magnitude transfers to our array, the affine component is bounded by 0.019–0.027 rad per spacing (Section 3, Supplement S4.2). At that level intervals a few degrees wide are safe, and the ambiguity only limits what can be certified below about 0.7–1 degree. The large-drift experiments are stress tests. A new exploratory analysis (Supplement S4.5) estimates the gauge without labels from the Doppler of static scatterers. It detects no change between eleven recordings of one day, but bounds it only at 0.031 in u, above the published scale, so it neither shows nor excludes drift. Measuring the affine component properly needs repeated captures of surveyed reflectors across sessions, which we have not done (Section 6).

The table below (Table 3 of the paper) shows where intervals of the narrow-trained ensembles, of width 0.030–0.036 in u, stop covering. "Uniform" averages over drift in [−B, B]; "worst fixed" is the lowest coverage at a drift held at ±B. The published bound corresponds to B between 0.019 and 0.027.

Drift bound B (rad/spacing)Uniform coverage (%)Worst fixed coverage (%)
0.01099.50–100.0099.02
0.02098.94–100.0096.84
0.03097.56–99.9488.68
0.05089.24–89.5215.06
0.07562.08–71.700.00
0.8006.16–7.680.00

Q2. Outside the narrowband far-field setting. With a spherical wavefront the gauge is approximate, not exact. Supplement S4.6 quantifies this: curvature separates paired worlds at close range and the planar approximation improves with distance. Non-affine channel errors (transmitter-only, receiver-only, random separable) do not translate the spectrum and are not invisible in our tests (Supplement S4.1). We did not study multipath.

Q3. Realistic multi-target perception models or real calibration drift. We have no such evidence. The learned experiments use a simulated single-source uniform linear array and small networks; the revision describes them as illustrations of the observation model, not a benchmark (Section 6).

Q4. When no law, bound or anchor is available. Then no finite-width coverage guarantee exists. This is the content of Propositions 1 and 2: the input cannot separate the worlds, and a set valid for a drift range of half-width B must have width at least 2B/κ. The paper does not solve that setting. It states what must be supplied, and how precisely, before a coverage claim can be made.

Novelty. We agree that the ambiguity is well known and that the paper carries it into uncertainty estimation. The revision presents the theory as a specialization and cites the works you list.

Real-data test of the anchor. On LiDAR-referenced recordings, nominal reference directions made anchoring worse at every anchor count tested, and after per-structure calibration the benefit was inconclusive (Section 4.3). We report this as a negative result.

Anchors KRMSE ratio, calibrated references [95% interval]RMSE ratio, nominal references
11.07 [0.78, 1.35]1.34
20.92 [0.67, 1.17]1.16
30.87 [0.63, 1.10]1.09
40.84 [0.61, 1.07]1.06

Ratios are relative to no anchor (below 1 means anchors help); six structures, one recording, 74 held-out pairs. The nominal-reference column is a post hoc analysis.

Reviewer_LF3SOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates gauge confounding between sensor calibration errors and target localization in coherent sensor arrays. The authors theoretically derive the associated identifiability limitations and lower bounds on predictive uncertainty, explore methods for restoring prediction coverage through nuisance parameter marginalization and external calibration information, and support their conclusions with synthetic experiments and controlled interventions on real radar ADC data.

Critical strengths

Meaningful research problem. The paper addresses the ambiguity between sensor calibration errors and target localization, which is relevant to the reliability of learning-based radar perception systems. The connection between physical identifiability and predictive uncertainty provides an interesting perspective on uncertainty estimation under distribution shifts.

Well-formulated theoretical analysis. The authors provide a mathematical formulation of gauge confounding and derive several theoretical results concerning representation invariance, observational equivalence, and prediction uncertainty. The analysis helps clarify the fundamental limitations of uncertainty estimation when different target configurations produce indistinguishable observations.

Multiple experimental perspectives. The paper combines analytical simulations, learning-based uncertainty estimation experiments, and controlled interventions on real radar ADC data. These experiments provide complementary evidence for the theoretical arguments and demonstrate the potential impact of calibration ambiguity on prediction reliability.

Reproducibility and discussion of limitations. The experimental protocols are documented in considerable detail, including numerical verification, multiple random seeds, and additional analyses in the appendices. The authors also explicitly acknowledge several limitations and report negative experimental results, which improves the transparency of the study.

Critical weaknesses

Limited novelty. The central ambiguity between phase calibration and direction estimation is a well-established problem in classical array processing. Similarly, the information loss caused by inappropriate invariant representations and the limitations of uncertainty estimation under indistinguishable observations are not entirely new. The main theoretical results largely extend established principles to a specific radar setting rather than provide fundamentally new insights.

Limited methodological contribution. The proposed uncertainty recovery relies heavily on known nuisance distributions or additional calibration information, which limits its practical applicability under unknown calibration shifts. Importantly, Appendix O shows that conventional direct prediction achieves nearly identical coverage and interval width when provided with the same calibration data. Therefore, the proposed framework does not demonstrate a clear methodological advantage over existing calibration approaches.

Insufficient experimental validation and generalizability. The learning experiments primarily rely on simplified synthetic single-source ULA data and small MLP models. Although real radar ADC measurements are included, they mainly involve artificially injected phase perturbations rather than naturally occurring calibration drift. The effectiveness of the proposed framework under realistic multi-target scenarios, multipath propagation, and unknown or time-varying calibration remains insufficiently validated.

Inadequate presentation and experimental comparisons. The overall organization and presentation of the paper could be substantially improved. A clear methodological workflow diagram is missing, making it difficult to understand the complete framework and its implementation. Furthermore, the experimental results are scattered across the main text and appendices, without sufficiently clear and systematic performance comparisons against representative existing methods. The effectiveness and practical advantages of the proposed approach are therefore not presented convincingly.

Questions for authors

First, the instructions for using your large model are missing; furthermore, the text is fragmented into too many minor points, lacking a cohesive focus. Second, compared to the substantive theoretical frameworks proposed by Wang et al. (2023) and Moussa (2026), does your theory offer any significant advancement—specifically regarding the prediction of lower bounds on risk? Third, I would like to know why increasing model capacity fails to resolve prediction ambiguity when the inputs are identical. Additionally, while the injection of phase ramps validates your theoretical predictions under controlled conditions, it does not necessarily prove that naturally occurring hardware drift exhibits the same characteristics. Fourth, both direct prediction heads and trajectory-based prediction heads can achieve satisfactory coverage when provided with the same perturbation distributions or calibration information. What, then, are the unique advantages of the trajectory-based approach? Moreover, in scenarios involving unknown drift or outdated calibration data, is it possible to clearly distinguish between theoretical guarantees, empirical coverage observed in controlled experiments, and actual reliability during real-world deployment?

1 comment

Thank you for the careful review and for noting the documentation of negative results.

Novelty. We agree. The phase–direction ambiguity is classical, and Theorem 1 and Proposition 1 specialize known principles. The revision says so in the introduction and in Section 5, and Appendix C1 compares our statements with the closest work one by one.

Advance over Wang et al. (2023) and Moussa (2026). None on lower bounds for risk, and we now say this explicitly (Section 5, Appendix C1). What we add is the consequence for interval width in direction units: an interval of width w can be valid at every drift only if the affine phase is known to within κw/2 (Section 2.4). Table 3 is consistent with this threshold.

Methodological contribution and the direct head. You are right that a direct head does as well as an orbit head when both receive the same current labelled panel; the differences are below 0.23% (Section 3, Supplement S3.3). With a supplied law the orbit readout improves the interval score at unseen bounds by 3.1% over a direct readout of the same network (Table S2). We do not claim an architectural advantage, and the revision states that no architecture is necessary. The paper is about which information supports which coverage claim, not about a new predictor.

Why more capacity does not help. When two worlds produce the same input, any function of the input returns the same output in both, whatever its capacity (Proposition 1). In our paired test of 5,000 cases every centre, interval and ensemble disagreement is identical, with coverage of 89.3–90.7% in one world and zero in the other.

Injected ramps versus natural drift. We agree. The injections verify the algebra and the implementation only. The revision says this in Section 3 and Section 6 and makes no claim about natural drift.

Guarantee, controlled coverage and deployment. Table 1 separates what each information source justifies. Under unknown drift or a stale panel there is no guarantee: with a stale panel both heads cover 5.8%. Section 6 and Appendix C3 list which statements are theorems under the model, which are simulation results, and which are exploratory real-data observations.

Workflow diagram and comparisons. Figure 2 gives the workflow. Table 12 (Appendix C2) compares all methods by the information they use, with coverage and width; an excerpt is in our general comment above.

Use of large language models. The statement is in Appendix C5.

We acknowledge that realistic multi-target scenes, multipath and time-varying calibration are not covered. That is stated as scope in Section 6.

Comments on the paper

1 comment

We thank the three reviewers. Their main points agree, and we accept most of them. The revision narrows the claims to what the evidence supports and adds one new exploratory experiment (item 7).

  1. Claims. The abstract and introduction no longer present the theory as new. The introduction states that the ambiguity and the bounds are known, and lists three contributions: the limits in the form of a precision requirement, a map from calibration information to the coverage claim it supports, and a test of reference transfer whose real-data outcome is largely negative.
  2. Main-text evidence. Section 3 is reorganized as one paragraph per row of Table 1. Table 2 (narrow versus wide training), Figure 3 (average versus fixed-drift coverage) and Table 3 (small-drift sensitivity) moved from the appendix to the main text.
  3. Practical scale. Section 3 now states the drift magnitude at which the analysis matters. At published phase-stability magnitudes, and assuming those transfer to the affine component, only intervals narrower than about 0.7–1 degree are affected. We state that we have no measurement of natural affine drift.
  4. Reference transfer. Section 4 lists the assumptions behind the anchor guarantee and what happens when each fails, and reports the real-data result as it is: nominal reference directions make anchoring worse, and after per-structure calibration the benefit is inconclusive and matched by a learned correction.
  5. Scope. Section 6 (Discussion and future work) pairs each open question with the experiment that would answer it.
  6. Length. Five exploratory appendices were reduced to rows in the table of claims and retained negative results (Appendix C3). The appendix was then cut to what the main text relies on: proofs, the protocols of the experiments reported in the main text, the comparison with the closest work, and the table of claims (14 pages, down from 39). The remaining experiments and controls are in a separate supplementary document, which the argument of the paper does not depend on.
  7. New experiment (Supplement S4.5). A label-free estimate of the affine gauge from the Doppler of static scatterers under platform motion, on twelve public recordings. The calibration idea is established practice and we claim nothing new for it. The estimate is reproducible (standard error about 0.006 in u per recording) and detects no change between eleven recordings of one day. Three of the five rules we fixed in advance failed, and the result bounds drift only at 0.031 in u. It does not show that natural drift occurs, or that it is absent at the published scale.
  8. Notation and wording. The shared-action index set is now Ξ, so that K denotes only the number of anchors. The full text was revised for clarity without changing any reported number.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.