acceptodds
Under review as a conference paper at ICLR 2027

Learning Approximate Isometric Embeddings for Efficient Template Placement

Abstract

The sphere covering problem has broad applications in science however, exact solutions are computationally infeasible on curved, high-dimensional manifolds because distances must be evaluated using a position-dependent metric. We present , a normalizing flow that learns an approximately isometric coordinate transformation, such that Euclidean distances in the latent space approximate metric-induced proper distances in the physical parameter space. This enables template placement using Euclidean distance checks in the latent space, avoiding repeated metric evaluations during placement. We evaluate in the context of gravitational wave template bank construction, where the sphere covering problem corresponds to the template placement problem. On a five-dimensional eccentric binary parameter space, reaches 97% coverage at a maximum mismatch of with 23% fewer templates than the state-of-the-art placement code , which reaches 94% coverage. Coverage is measured on independent injections drawn uniformly over the masses, spins and eccentricity, using the exact waveform match rather than the metric approximation. Training the flow takes about 2.5 h and placing the bank under 1 h on 32 CPU cores. will be made publicly available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Normalizing Flows, Representation Learning, Metric Learning, Geometric Deep Learning, Manifold Learning, Scientific Machine Learning, Gravitational-Wave Astronomy, Template Placement

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_6xqSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper discusses the problem of sphere covering, and concretely the matched filter template placement for gravitational wave detection. It proposes isobank, a method that trains a normalizing flow to transform signals into a Euclidean space that represents the proper distances in the physical parameter space. It uses a loss that combines global and local objectives. The method is applied to 3-dim and 5-dim GW signal spaces, where it showed improved coverage and mismatch compared with state-of-the-art method mbank.

Critical strengths
  • The proposed isobank builds on the idea of learning an invertible mapping into a representation space in order to construct a covering of the signal space. This is a clear motivation and reasoning, and produces performance improvement over state-of-the-art method (with a question to be answered below).
Critical weaknesses
  • The result and contribution of this paper is mostly confined to the field of GW detection, and methodological contribution is rather limited. There are some interesting technical points in the paper but never expanded upon. For example for the loss function which is a sum of 4 objectives, while the paper motivates each at an intuition level, it did not provide evidence or concrete analysis (e.g. through ablation) to examine the effect of each loss on the overall objective. The hand-tuned β\beta weights for these terms is another concern, as the paper did not discuss the sensitivity to them.
  • The key claim about performance improvement over mbank uses the default parameters, and it is not clearly explained to what extent that will undermine the claim. Since this is the centerpiece evaluation result, it should be presented with more clarity.
Questions for authors
  • In section 6.3 it's mentioned that mbank was run with its default settings. What other parameters are tunable for mbank, and how sensitive is performance to them? Could the performance be improved by adjusting its parameters, and how large of a performance gain would remain?
1 comment
AuthorsOpenReview ↗

Response to reviewer 3

Response to Reviewer 3

We thank the reviewer for their constructive feedback. We have uploaded a revised version of the manuscript, with all amendments highlighted in blue.

Response to comment 1 (methodological contribution, ablations and sensitivity analysis).

We agree with the reviewer that our work is currently focused on GW detection. However, a future avenue is to explore applications to waveform surrogates, where autoencoders have previously been used. The results in Appendix 8.5, comparing isobank with metric-preserving autoencoders, indicate that our method is more efficient for this task and could be extended to other sub-fields within GW physics. Applicability to other domains remains an interesting direction for future work.

We also agree that the individual contributions of the loss terms and the sensitivity to their weights deserved further investigation. We have therefore added an ablation study and sensitivity analysis in Section 6.3, with additional per-retraining results in Appendix 8.3.

The ablation study shows that LlocL_{loc}, LbndL_{bnd} and LedgeL_{edge} each measurably improve coverage in the eccentric space, whereas removing LglobL_{glob} has no detectable effect on coverage in either the eccentric or aligned-spin space. Training with LglobL_{glob} alone, however, results in substantially worse coverage. These results provide experimental evidence for the contributions of the loss components, while indicating that the global term could be omitted without measurable loss in the spaces considered here.

We also investigate sensitivity to the loss weights. The results are generally stable under factor-of-two changes to the coefficients of the local terms, although reducing the weight of LedgeL_{edge} or substantially increasing the tail-term weight (γ=10\gamma=10) degrades performance. This analysis provides a clearer understanding of the role of the different coefficients and identifies the edge term as particularly important for maintaining coverage.

We refer the reviewer to Section 6.3 and Appendices 8.3 and 8.5 for the full experimental results, as well as our response to Reviewer 2, Comment 2, for additional details on the ablation study.

Response to comment 2 (mbank parameter sensitivity and comparison).

The reviewer is correct that mbank may not be optimally tuned for our task. Another reviewer also pointed out that mbank may lack sufficient flexibility in our current setting. We therefore performed a sweep of minimal match and covering fraction with the default mbank model, as well as a brief fine-tuning experiment. Details are provided in Appendix 8.4. These experiments show that no tested mbank configuration achieves the same coverage with as few templates as isobank, demonstrating the efficiency of our method in the 5D eccentric space. We refer the reviewer to Appendix 8.4 for the detailed results.

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

This paper proposes to use learned approximately isometric embeddings for more efficient gravitational-wave template bank construction. They map the original parameter space to a latent space where metrics are easier to calculate. They train a normalizing flow to do this, and once this model is trained, they can use efficient Euclidean distance searches instead of the original physical metric. The authors evaluate on 3D and 5D parameter spaces, yielding modest improvements over naive methods.

Critical strengths
  1. Simple method: this method is concise and very intuitive. The loss involves a multi-scale loss that is intuitive
Critical weaknesses
  1. Limited comparisons: It's unclear whether the multi-scale loss that the author proposes is better than metric-preserving auto encoders, which have been proposed in the past. For me to raise my score, it would help to include comparisons to similar auto encoder baselines.
  2. Limited ablations: The method combines multiple intuitions for their multi-scale loss, but do not test what makes it work. It would help to separate out each component to isolate why each component is important experimentally.
Questions for authors

What is the end-to-end speedup over mbank? How many templates must be used in order for iso bank to be worth it to train, versus just using the naive method?

1 comment
AuthorsOpenReview ↗

Response to reviewer 2

Reviewer 2

We thank the reviewer for their constructive feedback. We have uploaded a revised version of the manuscript, with all amendments highlighted in blue.

Response to comment 1 (comparison with metric-preserving autoencoders).

We have reimplemented the method of Cabass et al. \citep{Cabass:2025wqg} for our setting and added a detailed comparison in Appendix 8.5. An important distinction is that their autoencoder learns a latent representation of the SVD coefficients of the templates, rather than mapping templates directly to their physical source parameters. In contrast, isobank is invertible by construction and maps between templates and source parameters. The SVD representation also limits the autoencoder's expressive power, particularly in the aligned-spin case, where it captures 93.1% of the space.

We therefore retain mbank as our main-text benchmark and discuss the autoencoder comparison in the appendix. We have clarified this methodological distinction in the main text. Our additional experiments (Appendices 8.5–8.6) show that the autoencoder can improve coverage by splitting the bank into multiple sub-banks, but this requires training a separate autoencoder for each sub-bank. Its cumulative fitting-factor distribution also has heavier tails than isobank, indicating less efficient template placement. We discuss how these differences may become more pronounced in higher-dimensional spaces, which we leave for future work.

Response to comment 2 (ablation study).

We have added an ablation study in Section 6.3, with per-retraining results in Appendix 8.3. Each variant retrains isobank from scratch with one loss term removed or one coefficient changed. Banks are compared at equal size (2.27×1062.27\times10^6 templates) and evaluated on the same injections.

Five retrainings of the full objective have miss rates of 2.75–3.73% (mean 3.36%). Removing LlocL_{loc}, LbndL_{bnd} or LedgeL_{edge} increases the mean miss rate by 1.5, 1.1 and 1.8 percentage points, respectively; using LglobL_{glob} alone more than doubles it. Removing the tail term (γ=0\gamma=0) increases the mean miss rate by 0.9 points. Removing LglobL_{glob}, however, has no measurable effect on coverage, also confirmed in the aligned-spin space. Varying the local-term coefficients by a factor of two generally produces results within the run-to-run spread, except for halving the LedgeL_{edge} weight. Setting γ=10\gamma=10 also degrades coverage. These results indicate that the local, boundary and edge terms are the main contributors to coverage in the tested spaces, while the global term can be omitted without measurable loss.

Response to comment 3 (end-to-end speedup and break-even point).

Appendix 8.6 now compares the computational costs of isobank, mbank and the autoencoder method, including preprocessing, training, placement and an estimate of the cost of searching one hour of data. This estimate is an order-of-magnitude approximation rather than a full matched-filtering benchmark.

For the comparison with mbank, the isobank bank contains 0.69 million (23%) fewer templates, saving approximately 320 CPU hours per hour of data searched. Its additional construction cost of 2.2 CPU hours is recovered after approximately 26 seconds of data. We emphasise that the main advantage of isobank is not faster bank construction, but more efficient template placement: achieving higher coverage with fewer templates reduces the subsequent computational cost of matched filtering.

Reviewer_6RHGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

Hidden layers are equivariant under U(1) or O(k); only the input and output maps break the symmetry. A large-width analysis of the U(1)-invariant activation magnitude has a transition at weight scale σ_W = 1. Above it, activations sit at a nonzero radius and a global phase is copied from layer to layer. The authors treat that phase as a Goldstone mode that carries signal without LayerNorm or skip connections. Equivariant RNNs and GRUs then improve long-delay copy, permuted MNIST, and the UEA archive.

Critical strengths

The loss concentrates capacity where coverage is decided: alpha in [0.6, 1.4], with a further term on the degenerate edge e >= 0.12. The audit uses the exact match on 10^4 injections, not the metric the flow was trained against. Held-out pairs in the 5D space fall along the diagonal, and constant-mismatch shells in the 3D space become nearly round after the map. In 5D the bank is 23% smaller (2.27M versus 2.95M), coverage rises from 94.17% +/- 0.2 to 96.64% +/- 0.2, and the worst injection moves from fitting factor 0.61 to 0.93. Training is about 2.5 hours and placement under an hour on 32 CPU cores, once pairs have been precomputed. The conclusion states the equal-dimension obstruction plainly.

Critical weaknesses

The abstract's 97% and 94% are rounded from 96.6% and 94.2%. In 3D, isobank uses 23,810 templates at 99.72% coverage against mbank's 23,167 at 99.80%. mbank is run at default settings for mu = 0.03, while isobank is allowed a coarser operating radius than r_target. A sweep of mbank's minimal match is required before the size reduction is attributed to the learned coordinates. The placement algorithm is still stochastic rejection sampling; the flow changes the distance, and metric evaluations move into pair generation rather than disappearing. Five aligned-spin dimensions, IMRPhenomD, and TaylorF2Ecc are a narrow test of a general covering method. Precession and higher modes, where the metric is much more expensive, are future work. Injection error bars are reported; flow retraining variance is not. The adaptive pair weights also make the objective depend on the current residual.

Questions for authors

See above.

3 comments
AuthorsOpenReview ↗

Response to reviewer 1

Response to Reviewer 1

We thank the reviewer for their constructive feedback. We have uploaded a revised version of the manuscript, with all amendments highlighted in blue.

Response to comment 1 (abstract percentages).

Thank you. We have amended the abstract to use the precise values (96.6% and 94.2%).

Response to comment 2 (3D comparison and mbank minimal-match sweep).

We agree that in the 3D case both methods perform similarly in terms of bank size. However, the placement speed-up is notable: approximately 2 hours for isobank compared with 10 hours for mbank's stochastic placement. We also compare the computational costs of both methods in a matched-filtering search setting in Appendix 8.6.

As requested, we performed a sweep of minimal match and covering fraction using the default mbank model, as well as a fine-tuning experiment. Details are provided in Appendix 8.4. These experiments show that no tested mbank configuration achieves the same coverage with as few templates as isobank in the 5D eccentric space.

Response to comment 3 (stochastic placement and metric evaluations).

We agree that \texttt{isobank} does not change the placement algorithm: placement remains stochastic rejection sampling, and the flow changes only the distance used. Metric evaluations are not eliminated but moved into pair generation, at a fixed cost of approximately 5×1045\times10^4 evaluations.

This is more than the approximately 2.5×1042.5\times10^4 evaluations used by mbank's random placement, but substantially fewer than those required by mbank's stochastic placement, which needs one evaluation per proposal. For a 2.95-million-template bank, this amounts to at least 3.3 million proposals, or approximately 1.8 days of computation. Once the flow is trained, placement requires no further metric evaluations. The additional cost of pair generation is offset by the efficiency of the resulting bank, as discussed in our response to Comment 2.

Response to comment 4 (generality and precessing systems).

We chose to keep the paper focused on eccentricity, which is of particular astrophysical interest because it provides insights into binary formation channels.

Nevertheless, we also tested isobank on a 4D precessing space using IMRPhenomXP. The parameter ranges were total mass 2525–100 M∘100\,M_\circ, mass ratio 11–55, primary spin magnitude 00–0.990.99, and tilt angle 00–π\pi, with a frequency range of 1515–10241024 Hz and the O3 PSD. The resulting bank contains 5,871 templates and covers 94.97±0.35%94.97 \pm 0.35\% of 4,000 injections at FF≥0.97\mathrm{FF}\geq 0.97. Without the edge term, a bank of 5,564 templates covers 92.67±0.41%92.67 \pm 0.41\%.

The computational advantage of our approach is particularly relevant in such spaces. In isobank, metric evaluations are required only for generating training pairs, at a fixed cost independent of bank size (approximately 15 minutes in this case). Placement then uses only latent distances. In stochastic placement, the metric is evaluated once per proposal, meaning that an expensive metric directly increases the total placement cost. We leave a more extensive investigation of these spaces to future work.

Response to comment 5 (flow retraining variance).

We have added retraining variance to the ablation study in Section 6.3. We retrained the flow five times in the eccentric space, using different initialization seeds and, in three cases, independently drawn training pairs. For a fixed bank size of 2.27×1062.27\times10^6 templates, the resulting banks miss between 2.75% and 3.73% of injections (mean 3.36%, standard deviation 0.38 percentage points), indicating stable coverage per template.

The latent distance scale varies between retrainings, however, so the radius required to obtain this bank size ranges from 1.311.31 to 1.82 rtarget1.82\,r_{target}. The radius therefore needs to be calibrated for each trained flow. In practice, we choose the largest radius whose bank reaches the target coverage on a small validation set. In the aligned-spin space, the required radius is more stable (0.910.91–1.05 rtarget1.05\,r_{target}), and five flows miss between 0.26% and 0.60% of injections.

Response to comment 6 (adaptive pair weights).

The weighting is not a separate rule applied at each training step. Rather, each ratio term is a fixed loss consisting of the mean error plus γ\gamma times the mean error over the worst-performing quantile. As training converges, this naturally focuses optimization on pairs that remain difficult. The apparent dependence on the current residual therefore arises from the gradient of this fixed loss. We have rewritten Eq. (10) to make this explicit.

The term improves coverage but is not finely tuned: removing it raises the mean miss rate by 0.9 percentage points, γ=1\gamma=1 and γ=3\gamma=3 give similar results, while γ=10\gamma=10 degrades the bank (Section 6.3 and Appendix 8.3).

Thanks for the authors responses, my evaluation of weak accept is unchanged. The response supplies the three missing checks I raised, so I raise soundness evaluation to 3. The abstract now says 96.6% and 94.2%. No tested mbank setting matches the 5D coverage at the same bank size. Five retrainings miss 2.75–3.73% of injections at 2.27 million templates. Pair generation is about 5×10^4 metric evaluations, against at least 3.3 million proposals for mbank’s stochastic placement. The authors also agree that placement itself remains stochastic rejection sampling.

The replies to the other two reviews point the same way. A reimplementation of the Cabass autoencoder is less efficient on this task. Ablations assign the coverage gain to the local, boundary, and edge terms; the global term can be dropped. The search-cost figure is an order-of-magnitude estimate: about 320 CPU hours saved per hour of data from 0.69 million fewer templates.

Two limits remain. The placement radius still has to be reset for each trained flow, from 1.31 to 1.82 times the target radius in the eccentric space. The new 4D precessing bank has no mbank comparison. Those keep the score at a weak accept.

AuthorsOpenReview ↗

Response to reviewer 1 additional comments

We thank the reviewer for their additional comments. We would like to clarify the two points raised:

  1. The reviewer is correct, and this is a genuine difference from methods that use the metric directly: in isobank the distance is learned, and its overall scale varies between retrainings on the eccentric space, so the radius yielding a bank of 2.27×1062.27\times10^6 templates ranges from 1.31 to 1.82 rtargetr_{\rm target}. The coverage obtained for that bank size is nevertheless stable (2.75–3.73% of injections missed across five retrainings), so the quality of the learned geometry is reproducible and only its scale must be fixed. On the aligned-spin space the radius is stable (0.98–1.04 rtargetr_{\rm target}). In practice the radius can be set with a few placements and a small injection set, each taking about ten minutes for the eccentric space; since any bank must anyway be validated by an injection campaign, this adds little to the overall workflow. The variation arises mostly along directions that the training pairs constrain weakly, and constraining these directions during training is a natural direction for future work.
  2. On the precessing space, we ran mbank's random placement with the Hessian metric at minimal matches 0.94–0.97, and placed isobank banks at four radii, all evaluated on the same 4,000 injections with exact matches. isobank reaches a given coverage with fewer templates: it covers 95.0% of injections with 5,871 templates, where mbank needs about 16,000 (16,991 templates give 95.4%), and 97.4% with 13,351 templates, where mbank needs 22,689 templates for 97.1%. mbank has the better worst case fitting factors (first percentile 0.95–0.97 against 0.91–0.94), as isobank's remaining misses concentrate at high spin. With mbank's default metric approximation, the bank at minimal match 0.97 has 2.0×1052.0\times10^5 templates (99.7% coverage).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.