acceptodds
Under review as a conference paper at ICLR 2027

Chirp-Modal Gaussian Mixtures for Joint Forecasting at Irregular Times

Abstract

Probabilistic forecasting of irregular time series demands coherent uncertainty across future times. Accurate pointwise intervals alone may neglect the temporal dependence that governs plausible trajectories. We propose chirp-modal Gaussian mixtures, which integrate damped oscillations with an explicit temporal covariance. Within each component, frequency and decay dynamically shift within the forecast window, enabling both the mean and the alignment of uncertainty to adapt over time. A commuting dynamical structure affords analytic means and transitions, while positive variance integration enables efficient joint sampling at irregular queries. We generate these components directly from history (CMD-X) or via latent reverse diffusion (CMD-D), and learn their temporal covariance with full-grid likelihoods. On all tested datasets, direct chirp mixtures consistently reduce energy and variogram scores relative to fixed-pole mixtures at comparable trainable parameter counts. At the headline configuration, iterative generation further improves joint scores, and same-checkpoint comparisons validate keeping mixture structure beyond a single Gaussian. Our approach combines adaptable temporal dependence with a choice of direct or iterative generation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Modeling, Continuous-time Models, Probabilistic Forecasting, Irregularly Sampled Data

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hRBuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes chirp-modal Gaussian mixtures for probabilistic forecasting of irregularly sampled time series, which extend fixed-pole modal representations by adopting time-varying frequency and decay. It also retains analytic state transitions and temporal covariance with two strategies: a direct one-pass mixture predictor (CMD-X) and a diffusion-based iterative predictor (CMD-D).

Empirical evaluations span four benchmarks, showing improvements in joint forecast metrics, energy score, and variogram score over fixed-pole variants and competitive baselines. Further controlled ablations, distributional analysis, and theoretical derivations support the empirical results.

Critical strengths
  1. Motivated formulation\textbf{Motivated formulation}. The combination of time-varying modal dynamics and explicit temporal covariance offers a principled way to model more natural future trajectories at irregular query times.
  2. Comprehensive empirical evaluations\textbf{Comprehensive empirical evaluations}. The experiments cover four datasets, multiple baselines, and 10 training seeds. Capacity-matched fixed-pole comparisons and controlled interventions help isolate the contributions of the claimed components.
  3. Flexibility at inference\textbf{Flexibility at inference}. The direct and diffusion-based variants offer different quality–efficiency trade-offs.
Critical weaknesses
  1. Insufficient validation of the latent independence assumption.\textbf{Insufficient validation of the latent independence assumption.} The Gaussian components assume independence across latent coordinates, which may restrict their ability to represent complex multivariate dependencies. Although cross-coordinate dependence can emerge through mixture assignments and nonlinear decoding, the paper does not empirically establish whether these mechanisms are sufficient.
  2. Novelty and positioning relative to LLapDiff\textbf{Novelty and positioning relative to LLapDiff}. The paper could handle the distinction from LLapDiff more carefully. Although LLapDiff utilizes time-invariant modal poles within each denoiser evaluation, these poles are already conditioned on the observation history and noisy latent state, allowing them to adapt across contexts and diffusion steps. Therefore, CMD's contribution is precisely introducing within-window temporal variation in frequency and decay, rather than adaptive modal parameterization itself.
  3. Limited empirical analysis of consistency across query grids.\textbf{Limited empirical analysis of consistency across query grids.} The covariance construction is guaranteed to be positive semidefinite on each requested grid, but the query-grid numerical approximation does not generally guarantee consistency across different query sets. Although the appendix provides a fixed-anchor alternative, its empirical implications remain unexplored.
  4. Incomplete characterization of quality–efficiency tradeoffs.\textbf{Incomplete characterization of quality–efficiency tradeoffs.} While the paper reports parameter counts and inference latency, a unified performance–latency Pareto analysis (plot) is missing, especially against external forecasting baselines.
Questions for authors
  1. The component-pool evaluation retained components that consequently contribute no samples, which makes it difficult to isolate the effect of mixture size from component ordering and sampling allocation. Please provide a cleaner comparison or give a self-consistent clarification.

  2. Analysis of marginal calibration across forecasting horizons seems limited; e.g., some results indicate noticeable undercoverage. For instance, on NOAA-UK, the empirical coverage of nominal 80% intervals is 62.4% for CMD-X and 73.1% for CMD-D. Please justify whether improvements in joint scores consistently translate into reliable marginal forecasts.

  3. How would the authors reposition and scope the paper given that the novelty differentiation is not sufficient, and perhaps a more precise comparison with competitive similar baselines would strengthen the contribution's positioning?

3 comments

Thank you for recognizing the formulation and controlled evaluation, and for identifying the analyses needed to strengthen the paper. We have added dependence and consistency diagnostics, clarified the component-pool experiment, and included the requested quality–latency plot.

1 Empirical and theoretical analysis of latent independence

Thank you for highlighting the need to examine whether within-component latent independence limits the model’s ability to capture multivariate dependence. In response, we have added a controlled empirical study and a theoretical characterization of the benefits and limitations of shared mixture assignment.

Controlled empirical evidence. Appendix H.4 tests shared mixture assignment directly using two-coordinate trajectories and identity encoder/decoder maps. A same-checkpoint product control preserves both full coordinate-trajectory marginals while removing cross-coordinate coupling. Retaining the learned coupling improves the joint log score by 0.0333 and 0.0941 nats/value at correlations 0.5 and 0.9, respectively (Table 28). Separately, comparing models trained with one and 25 components shows reductions in cross-covariance RMSE from 0.2523 to 0.1121 and 0.4541 to 0.2077, respectively (Table 29). These complementary comparisons demonstrate that shared mixture assignment can recover useful cross-coordinate dependence despite independence within individual components.

Theoretical characterization. Equation 54 gives a finite-mixture lower bound on population excess KL. At correlation 0.9, ten queries, and 25 components, the bound is 0.254 nats/value, compared with the fitted model’s empirical excess NLL of 0.332 nats/value. Thus, alongside the demonstrated dependence recovery, the analysis identifies a structural limitation of coordinate-factorized mixtures at a fixed context before decoding.

We agree that these controlled results do not establish whether shared mixture assignment and learned decoding adequately capture dependence on real datasets. Appendix A.5 additionally derives a shared-state covariance extension that permits within-component cross-coordinate dependence, providing a concrete formulation for further investigation. A controlled real-data comparison with this extension, using matched representations and training/evaluation settings, remains an important validation step. We acknowledge this remaining limitation and explicitly scope the conclusions in Appendices H.4 and M.

2 Precise contribution relative to LLapDiff

We agree with your distinction and now state it explicitly in "Introduction" and "Related work" sections: LLapDiff already adapts modal parameters to history, noisy latent state, and diffusion level. CMD introduces within-window variation in physical time, coupled to both the mean and temporal covariance.

CMD-X is the primary realization; CMD-D studies iterative generation of the same family. The controlled comparison is against fixed-pole direct mixtures sharing the representation, likelihood objective, and covariance construction. CMD-X lowers mean ES by 4.0–27.9% across all four datasets against closely capacity-matched controls (Table 1; Appendix J). This provides evidence for the precisely scoped contribution.

3 Fixed anchor consistency in practice

To address your concern, Appendix B.2 now evaluates fixed-anchor (FA) inference on five existing CMD-X checkpoints per dataset on PhysioNet and USHCN. We use 64 anchor intervals, retain the learned weights and decoder, and compare matched samples without retraining.

Cross-query consistency. We test subsets, off-anchor supersets, permutations and irregular close-pair grids, with 8,000 component cases per comparison per dataset. Maximum relative covariance discrepancies under the original query-grid construction reach 0.126 on PhysioNet and 0.587 on USHCN. FA reduces these to zero observed discrepancy in float64 and at most 1.74 × 10⁻⁶ in float32 with TF32 disabled, meeting the declared tolerances (Table 4).

Effect on forecasting. In a separate TF32-enabled evaluation, absolute mean ES/VS/CRPS changes across five paired checkpoints are below 0.02% of the original ten-seed reference means (Table 5). Repeated sampling gives similarly small aggregate changes. TF32-enabled consistency errors reach 1.45 × 10⁻³ and exceed the float32 tolerance, so predictive effects and numerical consistency are reported separately.

Guarantee and scope. Corollary 1 establishes cross-query consistency with history-only component parameters and weights, query-independent anchors and time conventions, and pointwise decoding. This covers CMD-X under these conditions; CMD-D’s query-dependent generator remains outside the guarantee. The main forecasting tables retain the original query-grid construction. The revision thus adds empirical consistency tests and a controlled assessment of predictive effects using existing checkpoints.

4 Quality and latency

Figure 9 in Appendix I now plots mean ES against reported latency for CMD and external baselines on USHCN and NOAA-UK. Both CMD variants are nondominated among the displayed configurations: CMD-X takes approximately 19–20 ms/window, while CMD-D takes 128–146 ms/window and attains lower mean ES. In these plotted settings, CMD-X also has lower ES and reported latency than LLapDiff and Chirp-MSE. Faster external methods occupy other frontier points. The plot makes the practical choice between direct and iterative generation explicit for these configurations.

5 Component pool allocation and the main results

To address your concern, we added a controlled comparison separating sample allocation from retained-component selection, using existing trained models and a fixed budget of 25 scored paths (Appendix D.1).

Cleaner comparison. We compare A, the original prefix/ordered allocation; B, balanced allocation over the same prefix; and C, balanced random code subsets for CMD-X. Thus B−A tests allocation at fixed membership, while C−B tests membership under balanced allocation. Every retained component contributes under B/C: at M=8, all eight receive 3–4 paths, whereas A uses only seven.

Results. At M=8, balancing lowers CMD-D mean ES by 1.49% on USHCN and 2.51% on PhysioNet, with VS also improving (Table 11). CMD-X mean ES decreases but VS increases. Subset effects depend on pool size and metric: random subsets lower PhysioNet CMD-X mean ES by 2.80% at M=8, but increase it at M=16 (Table 12). These controls directly quantify allocation and selection sensitivity.

Reliability and scope. We use five checkpoints per model and dataset, averaging five matched evaluation repetitions within each checkpoint before reporting paired changes and across-checkpoint SDs. Table 13 separately reports evaluation variability. At the primary M=S=25 setting in Tables 1–3, every component supplies one path and the allocation protocols coincide, so the primary comparisons have no unused components. We therefore interpret intermediate curves as inference-time retained-pool operating points, rather than isolated effects of training mixture size.

6 Marginal calibration in context

Thank you for drawing attention to this distinction. We agree that better joint scores do not guarantee nominal marginal coverage, and that the NOAA-UK results reveal a remaining calibration limitation. For nominal 80% intervals, CMD-X and CMD-D achieve empirical coverage of 62.4% and 73.1%, respectively, both below the target.

The full results provide additional context. Empirical coverage of nominal 80% intervals is 79.0%/82.0% on PhysioNet and 74.4%/76.2% on USHCN for CMD-X/CMD-D (Tables 17–18). On NOAA-UK, CMD-D’s 73.1% coverage improves substantially over 53.5% for LLapDiff and 51.6% for ProFITi, alongside lower CRPS (Table 20). These comparisons show improvements in forecast scores and coverage relative to those baselines, while leaving residual undercoverage.

Calibration is also not uniformly strongest across methods: TFMixer achieves 73.6% coverage and lower PIT-ECE on NOAA-UK, despite worse ES and CRPS. We therefore distinguish improvements in predictive scores from evidence of calibrated uncertainty.

We recognize that these aggregate diagnostics do not fully address your concern about calibration across forecasting horizons: pooled coverage can conceal horizon-specific deviations. Appendix E reports coverage together with interval width, but a horizon-resolved evaluation of these quantities remains an important validation step. We have explicitly acknowledged the remaining calibration limitation and scoped our claims to the demonstrated forecast-score improvements, without presenting them as proof of nominal coverage across horizons.

We hope these targeted experiments, theoretical characterization, and clearer scope address your concerns and support a more favorable assessment of the revised paper.

Reviewer_hRBuOpenReview ↗

Re: Official Comment by Authors

Thank you for the response and the substantial additional experiments and analyses. The revision addresses most of my original concerns and improves the paper.

In particular, the distinction from LLapDiff is now much clearer, and the capacity-matched fixed-pole comparisons provide more convincing evidence for the contribution of within-window frequency and decay variation. The additional cross-coordinate dependence experiment, fixed-anchor consistency analysis, and component-allocation controls also strengthen the methodological and empirical evaluation. The quality–latency comparison further clarifies the practical tradeoff between CMD-X and CMD-D.

A few limitations remain. First, the cross-coordinate dependence experiment demonstrates useful but incomplete recovery under a controlled synthetic setting, without establishing its sufficiency on real datasets. Second, the fixed-anchor consistency guarantee applies to CMD-X under specific conditions and does not extend to CMD-D; numerical discrepancies under TF32 also remain. Third, marginal calibration, particularly across different forecasting horizons, remains insufficiently characterized, although the authors now acknowledge this limitation and appropriately scope their claims.

Overall, I appreciate the authors' substantial efforts. I am therefore inclined to revise my original assessment more favorably, yet there is not an option between 2 and 3. Although the rating is not changed, I am inclined towards a borderline instead of a solid 3.

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

This paper introduces chirp-modal Gaussian mixtures for probalistic irregular time series forecasting. The architecture models the developement of latent variables over time by a Gaussian mixture Process that is defined by chirps. How to compute the mean vector and covariance matrices from the chirps is derived in detail. The inference of the chirp-parameters is however only briefly explained and in some important details (espcially the model's encoder) are fully absent from the paper. The chrip induced Gaussian Mixture models the latent state of forecasting targets, which stem from a pre-trained and then frozen VAE. The VAE is not explained in furhter detail. For instance, it is unclear if the VAE operates on single query/target time points or on the matrix of all possible targets. The authors also mention that the encoder (history summarizer) is frozen when the predictor is trained. However, it is not clear when and how it is trained then.

The paper introduces two models which use the chirp-model Gaussian mixtures. One mode, CMD-C, predicts the predicts the chirp paramters from the observed history and minimilizes the NLL of the respecitve Gaussian mixutre model and the target hidden states which stem from the VAE. The other model CMD-D, uses diffusion type training to predict the latent variables from the VAE by de-noising.

In the experiments the model is compared with LTI, which uses fixed poles, but is not furhter described. Additionally, authors compare with recent joint models for IMTS forecasting. The models are evaluated using CRPS, VS and ES. The Chirp based models are outperforming the baselines consistently on these datasets.

Critical strengths

The paper carefully derives how to compute the mean vector and covariance matrix from the chirp model. The experimental results are impressive.

Critical weaknesses

First, the paper lacks some important details about the encoder, VAE and the LTI ablation. Second, the paper does not utilize established IMTS benchmarks from the ProFITi and MOSES paper. Justifying why these experiments were not appropiate for this work, or better evaluating CMD-C/CMD-D on these tasks in order to comapre with published results would make the comparison more trustable.

Questions for authors
  • Could give the implementational details as mentioned in the weakness section ?
  • Could you compare with published results (e.g. from the MOSES paper) ?
2 comments

Thank you for your positive assessment of the derivations and experimental results. We have addressed your two requests with explicit implementation specifications and a new comparison against published benchmark results.

1. Why the main paper uses h∈12,168h \in 12, 168 and new evaluation on the official MOSES USHCN benchmark

  • Motivation for the main benchmark protocol: In the official MOSES/ProFITi USHCN protocol, the target window consists of only the next 3 observation rows (h=3h=3) after t≤0.75t \le 0.75. Because a central goal of CMD is modeling within-window frequency/decay drift and multi-step trajectory covariance, our main protocol evaluates longer horizons h=12h=12 queries on USHCN and PhysioNet (following LLapDiff's PhysioNet setup) and h=168h=168 hours on NOAA-UK and BMS-air, while training and tuning MOSES, ProFITi, and eight other baselines across all four datasets (Table 3; Appendix K Tables 32–33).
  • New comparison on the official MOSES USHCN splits and published numbers: Appendix D.2, Tables 14–15, evaluates CMD-X on the official five station partitions and three-row target windows, using the benchmark's metric implementation, aggregation, and source-data scale. This adds a held-out-station task to the main chronological USHCN evaluation with its 12-query horizon.
ModelCMD samplesES ↓CRPS ↓MSE ↓
MOSES published—0.552 ± 0.0440.220 ± 0.0190.411 ± 0.099
ProFITi reported in MOSES—0.452 ± 0.0440.182 ± 0.0070.308 ± 0.061
CMD-X1000.464 ± 0.0480.190 ± 0.0180.279 ± 0.090
CMD-X1,0000.459 ± 0.0510.188 ± 0.0190.275 ± 0.089

At 1,000 samples, CMD-X's reported means are 16.8% lower in ES, 14.5% lower in CRPS, and 33.1% lower in MSE than MOSES. Compared with ProFITi, its ES and CRPS are 0.007 and 0.006 higher, while its MSE is 0.033 lower. The additional task therefore provides concrete evidence of competitive performance beyond our main protocol.

CMD results are mean ± SD across five splits with one training seed per split; published baseline conventions are retained. Appendix D.2 documents the context/training adaptations and selection limitation: split-0 validation was used to choose the shared configuration, and some of those stations occur in later test splits. We therefore qualify cross-split independence and do not infer statistical equivalence or a fully matched comparison from these summaries.

2 VAE history summarizer and LTI implementation

The VAE processes each target timestamp separately across entities, using a 3-layer encoder and decoder, each with four heads, width 128, and feedforward width 256. Latent width is 16, or 24 on BMS-air. Pretraining uses masked reconstruction MSE and KL regularization, plus view consistency on USHCN/PhysioNet, for up to 600 epochs. The forecasting targets are standardized posterior means; decoding is pointwise in time.

The history summarizer combines 64-channel temporal convolutions, a two-layer/four-head temporal Transformer, and learned-query pooling. It is pretrained as a history-reconstruction autoencoder for up to 200 epochs. Both modules are randomly initialized, pretrained independently on training data, selected by validation objectives, and frozen for CMD training (Appendix C.1, Table 9).

The LTI control retains CMD-X's stochastic Gaussian mixture, representations, and full-grid mixture-NLL objective, but fixes frequency and decay within each forecast window. These parameters still depend on history and component code. Table 1 includes equal-mode and closely parameter-matched comparisons; Appendix J, Table 30, specifies what is held fixed.

Given your positive assessment of the derivations and results, we would appreciate reconsideration in light of the added benchmark evidence and the implementation details that were missing from the submission.

I thank the authors for their response. After reading the remaining reviews and rebuttals and re-reading the improved paper I decided to keep my score. Additionally to the limitations raised by other reviewers I am concerned with the hyperparameters space of baselines such as ProFITi, (which appears to be the most important baseline), as the search space differs significantly from the experiments of the ProFITi paper.

Reviewer_m1o5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the problem of joint probabilistic forecasting under irregular time-series. The paper proposes to address this problem by introducing chirp-modal gaussian mixtures, where temporal variances are handled within each component. The designed method, CMD, outperformed previous joint probabilistic forecasting methods, with ablations on model architectures.

Critical strengths
Critical weaknesses

The introduction is poorly written and undermines the paper's intellectual contribution. The introduction does not position CMD against previous joint irregular-forecasting well. The third paragraph acknowledged that previous works provided solutions to joint irregular-forecasting but didn’t distinguish what they have missed and what CMD is addressing. And terms like “joint forecast” as well as “history-conditioned component family” are appearing before they’re defined in Section 2.1. Furthermore, the authors propose two solutions, CMD-X and CMD-D, but do not commit to either as the main contribution. All of these make the section hard to read and undermine the intellectual contribution of this paper. 2. The experiments are missing details and do not align with the claims. The paper introduction and abstract highlights within-component temporal covariances are needed to model the sequence as detailed in section 2.3-2.4. Yet, the block-diagonal ablation, as shown in Figure 4, barely hurts the ES on USHCN and PhysioNet. The major performance gain is therefore from drifts and the mixture structure, since gains shown in Tables 1 and 10-13 are significantly larger.

Questions for authors

Please see the weaknesses and the following missing details:

What are the “student-t”, “diagonal Gaussian”, and “low-rank Gaussian” in Table 10-13 established? The authors mentioned them as distributional controls but didn’t provide any details on their implementation. How are the VAE and history summarizer initialized? I can only see in Appendix C that both are frozen, and Appendix K seems to say there is a pretraining stage before stage-2, but no details are provided How are the baselines configured and tuned? The authors only mentioned that they performed parameter search, but without search space specifications and finalized configurations, except for some learning rates.

2 comments

Thank you for identifying where the presentation obscured the contribution and experimental evidence. We have substantially revised the positioning and implementation description. Below we clarify what CMD contributes and what the controlled results establish.

1 Contribution and positioning

Our main contribution is the chirp-modal Gaussian component family: frequency and decay evolve within the forecast window and jointly determine the mean and temporal covariance. Section 1 now defines joint forecast immediately, credits ProFITi and MOSES, and explicitly acknowledges LLapDiff's existing adaptation to history and diffusion state. CMD adds variation within physical forecast time during each component evaluation.

We now commit to CMD-X as the primary direct realization, with CMD-D as an iterative extension. The evidence for this contribution is concrete: against capacity-matched fixed-pole mixtures, CMD-X reduces mean ES by 4.0–27.9% and VS by 3.8–42.0% across all four datasets (Table 1). Both models retain temporal covariance and share representations and the mixture-likelihood objective, so these comparisons assess evolving dynamics within a common design.

2 Evidence for within component temporal covariance

Your observation about the small ES changes on USHCN and PhysioNet is correct. The complete intervention results also show a consistent increase in mean VS and larger ES effects on the environmental datasets:

DatasetES increase after removing temporal covarianceVS increase
NOAA-UK5.48%9.82%
USHCN0.61%8.10%
BMS-air4.32%8.07%
PhysioNet0.36%5.30%

This intervention preserves each per-time distribution and the dependence induced by sharing a component identity throughout a path. It removes only residual cross-time covariance within components. VS assesses pairwise trajectory differences, providing a complementary diagnostic of that dependence (Section 3.3; Figure 4; Appendix G).

The separate synthetic study provides a controlled likelihood check: removing temporal covariance raises excess joint NLL from 0.00255 to 0.31314 nats/query while preserving marginal NLL. Collapsing the mixture to a Gaussian with the same full mean and covariance raises excess joint NLL to 0.10040 (Appendix H.3, Table 26).

Together, these results support evolving dynamics, mixture shape, and residual temporal dependence through distinct tests. We have revised the abstract and discussion to avoid claiming that covariance explains most overall gains. The experiments do not provide an additive attribution, and an inference-time intervention does not measure performance after retraining without covariance.

3 Distributional controls

Appendix C.2 and Table 10 now specify the controls over flattened future latents: independent diagonal Gaussian; low-rank Gaussian with Σ = FFᵀ + diag(s²); and independent Student-t with ν ≥ 2.1. Each keeps the VAE frozen while jointly fine-tuning the pretrained history summarizer and output head. The appendix supplies their likelihood objectives, sampling, and numerical safeguards; Table 31 supplies search settings.

4 Initialization and pretraining

Both modules are randomly initialized, independently pretrained on the training split, and then frozen for CMD-X/D. The VAE uses a 3-layer, 4-head Transformer encoder/decoder of width 128, operating across entities separately at each target time. It uses masked reconstruction and KL regularization; forecasting targets standardized posterior means. The history summarizer combines 64-channel temporal convolutions, a 2-layer, 4-head temporal Transformer, and learned-query pooling, pretrained to reconstruct history. Appendix C.1 and Table 9 specify the remaining architecture, dataset-specific losses, optimization, and checkpoint selection.

5 Baseline tuning

Appendix K and Tables 31–34 expand the search grids, fixed settings, selection protocol, model sizes and and dataset-specific selected values (annotated per dataset in Tables 31–33). Each searched baseline uses 8–12 seed-0 search trials, up to 100 epochs with validation-based early stopping, followed by separate final runs over ten seeds, each with a 200-epoch limit.

We hope the clearer contribution, controlled evidence, and expanded specifications address the central grounds for your assessment, and would appreciate your reconsideration of the revised paper.

I have read the authors response and prefer 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.