acceptodds
Under review as a conference paper at ICLR 2027

Re-Excitable Response Representations from a Single-Waveform Radar Observation

Abstract

Pulsed radar measurements are intrinsically conditioned on the transmitted waveform, and the same scene can therefore produce substantially different observations under different excitation conditions. We study radar re-excitation: predicting how a scene observed once under a source waveform can be measured under another known waveform, without physical re-acquisition. The central challenge is that the source may not adequately capture response directions that become important under the query waveform, so source consistency alone is insufficient for reliable prediction. We address this ambiguity with a Re-Excitable Response Representation constructed from a single source observation. The representation anchors measurement-supported content to a source-evidence response, organizes unresolved content through shared propagation trajectories and compact event-aligned responses, and performs observability-conditioned closed-form completion with structural intervention restricted to source-weak directions. The completed response can subsequently be re-excited under known waveform queries without requiring target-waveform scene observations. Across comprehensive synthetic benchmarks and real-world tests, our method achieves the lowest MSE and highest SSIM among six physics-guided and learning-based baselines. Re-excited measurements further enable two frozen inversion models to approach native-waveform performance, demonstrating a reusable interface between varying sensing waveforms and downstream physical inference.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Physics-Informed Learning, Domain Adaptation, High-Resolution, Electromagnetic Inversion

Reviews

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

This paper studies radar re-excitation, i.e., predicting how a scene observed once under a source waveform would be measured under another known query waveform, without physically reacquiring the scene. The key observation is that source consistency alone is insufficient for cross-waveform prediction because response directions that are weakly observable under the source waveform may become important under the query waveform. The paper formalizes this issue through cross-operator observability transfer and proposes a Re-Excitable Response Representation that can be reused across different waveform queries. Methodologically, the proposed representation combines three main components: (1) a source-evidence anchor that preserves response content supported by the observed source measurement; (2) an event-aligned structured continuation that models unresolved response components using shared propagation trajectories and compact local response bases; and (3) a source-observability-gated completion that restricts structural correction primarily to source-weak response directions. The resulting response state is query-independent and can subsequently be projected through different known query operators. The method is validated on seven synthetic cross-waveform super-resolution settings covering frequency, geometry, heterogeneous-medium, and noise shifts, as well as the real-world TU1208 200→400 MHz dataset. The paper further evaluates downstream cross-waveform inversion using frozen segmentation and permittivity-reconstruction models, showing that re-excited measurements can approach native-waveform performance without retraining the downstream models. Component ablations are also provided to support the roles of the evidence anchor, structured continuation, and observability gating.

Critical strengths
  1. The paper addresses a practically meaningful and well-motivated problem. The paper provides a potential interface for reusing existing radar measurements and frozen downstream inversion models across different waveform configurations, thereby reducing the need for repeated data acquisition and waveform-specific retraining. This practical motivation is particularly compelling because waveform mismatch is a genuine obstacle to transferring sensing and inversion pipelines across acquisition conditions.

  2. The method has a coherent connection between the theoretical motivation and algorithmic design. The paper first identifies cross-operator observability transfer as the key ambiguity, and then explicitly separates source-supported response content from structurally inferred content. The observability-gated composition is particularly well aligned with this motivation, since structural intervention is restricted to directions that are weakly supported by the source observation rather than modifying the entire response indiscriminately. The ablation in which removing structured continuation preserves near-zero source error but causes a substantial degradation in target-domain reconstruction provides useful empirical support for the central claim that source consistency does not imply query accuracy.

  3. The experimental validation is broad and includes downstream utility rather than only signal-level similarity. In addition to synthetic waveform conversion, the paper evaluates real measured data and two frozen downstream inversion tasks. The latter is important because it demonstrates that the converted measurements can serve as a practical interface between varying sensing waveforms and existing inference models, rather than merely improving MSE/SSIM at the measurement level.

Critical weaknesses
  1. The practical applicability depends on relatively strong calibration and structural assumptions. The method requires calibrated source/query waveforms and relies heavily on event-aligned propagation trajectories and compact local response models to continue information into source-weak directions. Therefore, its success appears contingent on the validity of these structural assumptions. It is not yet clear how robust the method would remain under more complex real-world conditions involving inaccurate waveform calibration, complicated antenna responses, strong multiple scattering, or propagation patterns that deviate from the assumed event structure. The paper itself acknowledges several of these limitations.

  2. The comparison with learning-based baselines may not fully establish superiority under a balanced evaluation protocol. The learning-based baselines are trained only on 3,000 paired G15→G25 examples, but are then evaluated on substantially different frequency gaps, geometries, media, and noise conditions. Their poor performance under G9/G5→G25 and other out-of-distribution settings is therefore expected to some extent. Since the proposed method is explicitly designed to exploit known waveform physics and source-side structure without paired training, this protocol strongly favors its intended generalization regime. Additional comparisons with learning-based methods trained or adapted across multiple waveform conditions would provide a more convincing assessment of the proposed method's advantage.

  3. The real-world evidence is still limited relative to the generality of the claims. Although the inclusion of TU1208 is valuable, the main quantitative evidence and all downstream inversion experiments remain dominated by controlled synthetic settings. Given that the method is motivated as a reusable cross-waveform interface for practical radar systems, validation across more real datasets, sensing configurations, and acquisition environments would be important to establish that the proposed representation generalizes beyond the specific calibrated settings studied here.

Questions for authors

See weaknesses above.

Reviewer_sHb3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies radar re-excitation, which aims to predict radar measurements under a different known waveform from a single source-waveform observation. The authors point out that source consistency alone is insufficient for reliable prediction. The authors then propose a re-excitable response representation that combines a source-evidence anchor with event-aligned structured continuation through an observability-gated closed-form composition to enable cross-waveform measurement prediction. The method is evaluated on EM super-resolution and cross-waveform inversion tasks using synthetic and real-world data, achieving strong performance compared with physics-guided and learning-based baselines.

Critical strengths

Physically motivated and well-structured approach: To address the problem that source consistency alone does not guarantee query accuracy, the paper combines a source-evidence anchor with event-aligned structured continuation and uses observability gating to control the completion process. This approach is supported by the experimental results and ablation studies.

Critical weaknesses

Insufficient comparison with learning-based baselines: All learning-based baselines are trained only on paired G15→G25 data but are evaluated under unseen frequency conditions such as G9→G25 and G5→G25. Although this setting can evaluate out-of-distribution generalization, the lack of comparisons with models adapted or trained for other frequency conditions may limit the performance of the learning-based baselines, thereby weakening the evidence for the proposed method's performance advantages over data-driven approaches.

Questions for authors

Can the authors provide or discuss the performance of learning-based baselines trained or adapted for other frequency configurations, such as G9→G25 and G5→G25? While the advantage of the proposed method in not requiring paired cross-waveform training data is appreciated, it would be valuable to understand the performance gap between the proposed method and learning-based baselines trained or adapted for the corresponding frequency configurations. Such comparisons would help better assess the advantages of the proposed method itself and the impact of the training–test mismatch on the reported results.

Reviewer_y9izOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The problem considered in this paper is radar reconstruction using re-excitation, i.e. sending two different waveforms at the same target. The proposed method is based entirely on linear algebra and Fourier analysis and as far as I can see no learning is involved. The results compares favourably with several learning based approaches for both simulated and real data with known ground truth.

Critical strengths

Sound derivations. Thorough testing against baselines.

Critical weaknesses

Not critical per se. But perhaps the application would be better to present in a more specialised venue especially given that the learning aspect is limited.

Questions for authors

More clear motivation why this work belongs a ICLR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.