Thank you very much for the detailed and critical evaluation of our paper and the precise questions!

Response to Weaknesses

Reviewer: Positional encoding for [...] manifolds, such as the sphere, have been considered previously in [...] Koestler et al. (2022), which is not cited but should be.

Thank you for the valuable insights in Koestler et al. (2022). We modified the introduction to acknowledge this work (blue font).

Reviewer: Limitation to low degrees L.

While the computational complexity increases with higher L, as shown in Figure 5, we would like to highlight that the computational runtime is not excessive even until L=100, which is the largest L harmonic function we pre-computed. Pre-computing the analytic functions of higher L is easily possible, but we found that such a high L was not necessary for the experiments of this paper.

To illustrate that runtime is not a major limitation, a forward pass of 256 coordinates up to L=100 (in total L$^2$ harmonics) takes only 716 ms ± 17.7 ms on a Macbook CPU. In comparison, a forward pass through a ResNet 50 vision model on a [256 x 3 x 224 x 224] image batch takes 5.33 s ± 225 ms on the same computer. Hence, a SirenNet(SH) location encoder (even with a high L) will likely still contribute less to the runtime than a common vision model.

Reviewer: It is indeed the case that combining the spherical harmonic encoding with SIRENS typically gives the best performance (see Table 1). However, in many cases the improvement is not that substantial.

While the performance gain of SH may seem comparatively small, it is generally consistent across all datasets and across neural networks. We show in Fig 4b) that SHs are most effective at high latitudes and polar areas where the spherical geometry of our planet matters the most. Here, other DFS-based location encoders systematically lose accuracy since their underlying assumptions on the geometry (i.e., the rectangular domain of lon/lat) are not met. Since only a few test points are located at high latitudes or polar areas, the accuracy benefits are numerically small in the tables.

Response to Questions

Reviewer: Could the authors elaborate on how their work fits into the context of prior work by Grattarola & Vandergheynst (2022) and Koestler et al. (2022).

We see our work targeted towards the concrete modality of geographic location encoding in lon/lat coordinates on our spherical Earth. This is reflected by the tested datasets (land-ocean classification, ERA-5 climate science, iNaturalist species distribution modeling). We understand this work as a special case of learning implicit neural representations on the Sphere as specific non-euclidean geometry. We highly appreciate the fundamental works of Koestler et al. (2022) and Grattarola & Vandergheynst (2022). The generality of these papers to learn functions on arbitrary geometries through the expression as eigenvectors in graphs is highly valued. In this work, we only focus on the sphere of our Planet. Hence, we can focus on purpose-built, efficient implementations of Spherical Harmonics with analytically pre-defined functions of certain degrees and orders.

Reviewer: It seems combined spherical harmonic encoding with SIRENs is generally best. However, for the ERA 5 dataset it seems a fully connected network was significantly better than combining with SIREN. Do the authors have any idea why a fully connected network was best here?

FcNet is a well-tested network for location encoding. In our opinion, FcNet remains a reasonable choice when it is paired with a suitable positional encoding. SirenNet achieves similar and often better results and, crucially, SirenNet is less sensitive to the underlying positional encoding and predicts reasonably well even with direct encoding of longitude and latitude. In Section 2.3, we have shown that a specifically initialized SirenNet layer becomes equivalent to the Grid DFS-embedding. Out of these reasons, we see SirenNet as an architecturally simpler replacement over FcNet that archives reliably good results irrespectively of the positional embedding.

Reviewer: Do the authors have any thoughts on how they could extend the method to higher degrees L?

One heuristic to enable higher degrees L would be to pre-compute the numerical values of harmonics of explicit training points with a high L (like L=1000). These harmonics can then be loaded directly from the disk.

A more principled solution is likely found in Geodesy, where the Earth's gravity field is stored as weighted spherical harmonics up to several hundred Legendre polynomials (i.e., Linear(SH)). In this field, Jekeli et al., 2006 and Balmino et al., 2012 investigated approximations that allow integrating signals with very large L. We plan to explore this direction in future collaborative work.

Thank you for the detailed evaluation, precise questions, and insightful comments!

Response to the Weaknesses

Reviewer: It would have been interesting to see more comparisons on datasets [...], e.g., the fMoW dataset.

We considered including FMoW initially but decided against it, as the accuracy benefit of adding coordinate information is only 1.62 % (see Mai et al., 2023; Sphere2Vec Table 3; no std devs given). We believe that this margin is too small to show any systematic differences between the location encoders. Instead, we focused rather on iNaturalist, where the benefit of adding geolocation adds a 12.3% improvement (Table 1c) over the image-only model.

Reviewer: Would it be possible to integrate this method other tasks more common with ERA5 such as downscaling?

The location encoder is not restricted to gridded data and can predict ERA5 variables continuously. Hence, generating a 1km rather than 30km grid (as in downscaling) is possible on a technical level. Still, adding further (higher-res.) covariates, like detailed topographic features, is likely necessary to improve the effective resolution. The location encoder can be used here as a module for a larger downscaling model.

Response to Questions

Reviewer: Intuition about the combination of SirenNet and some DFS based encodings performing worse than SirenNet alone?

Sphere{C,M} keep one lat/lon coordinate at the original (lowest-frequency) scale while scaling up the other. We hypothize that the lowest scale may be too low-resolution, which can be potentially detrimental. Beyond that, we have no further intuition other than that DFS-embeddings are not optimal for the Sphere in general (violated assumption of rectangular lat/lon domain). Regarding the good performance of SirenNet alone (i.e., with Direct), we show in Section 2.3 that a single Siren layer with specifically set weights can produce the same embedding as Grid. Hence, it is not surprising that Siren with direct coordinates performs competitively well.

Reviewer: I may have missed this information but what is the number of Legendre Polynomials in the experiments results reported in the tables?

We performed hyperparameter tuning separately for each NN-PE combination, as described in the Appendix Section "A.1 Implementation details and hyperparameter tuning".

Concretely, every model in the results tables has a different configuration for the SH models. We tune the Legendre polynomials L between 10 and 30 in steps of 5 polynomials. In the best hyperparameter file, we find the entire value range (10 to 30) present. Indeed, L controls the smoothness of the interpolation directly. A too-low L may be too low-frequency to capture the signal, while a too-high L may lead to overfitting around particular training points and not considering nearby test points.

Reviewer: I would also be curious to have the performance of the different models of Figure 5 with the number of polynomials alongside the computational efficiency.

Thank you for this suggestion. We provide the accuracies of the models below. Note that SphereC+ does not have an L parameter to control the degree of interpolation, but an $r_{min}$ parameter that determines the minimum radius. We compare both using the maximum resolvably frequency formula $r_{min} = f_{max} = \frac{180°}{2L}$.

You can find the accuracies of the runtime experiment for Figure 5 below:

Note that SphereC+ is not as accurate as in the comparison tables, as $r_{min}=4°$ is likely too restrictive (many, but not all, optimal $r_{min}$ hyperparameters are 1). We chose $r_{min}$ to be comparable to L rather than optimal with respect to predictive accuracy (as in Tables 1,2). Accuracy-wise, Siren(SH) with L=20 is slightly more accurate than Siren(SH) L=40. Hence, a higher L does not automatically lead to better accuracies.

Reviewer: What is your take on SphereM poor performance?

Thank you for noticing that. We double-checked our implementation of SphereM{+}, which corresponds exactly to the code provided by Mai et al., 2023.

To make sure, we also re-ran the land-ocean classification on a different computer (now Macbook Pro on CPU/originally a GPU workstation), which resulted in very similar results to Table 1b:

Hence, we do not have an explanation for the poor SphereM performance and can only stress that we run all models in the same environment.

Thank you for the precise and accurate analysis of our work!

Reviewer: The paper is quite dense, with prevalence of abbreviations and mathematical notation over the description, which makes it difficult to read, especially in sec.3 Datasets and experiments. Readers without the knowledge of the specific datasets won't be able to understand the application. I would suggest reduce the number of the use-cases to 3 (g.g ERA5 and iNaturalist) and describe in details what exactly did authors do.

We greatly appreciate your comments regarding the density and details and have taken your suggestion by extending the dataset section 3 in the revised version by adding additional sentences marked in blue in the revised paper.

We would still favor keeping all four use-cases/datasets in the main paper, as every dataset reflects a different aspect of performance. We did our best to line out the major goals and findings of our study with a clear overall structure and text.

Thank you again for the accurate summary of for paper and we remain at your disposal if any further questions remain.

Thank you very much for your careful and critical reading of our work.

Response to Main Questions

Reviewer: When are the performance gains from SH worth it?

The quantitative differences between SH-based encoders and DFS-based encoders may be numerically small but are systematic for a) different neural networks and b) present across different datasets (Tables 1 and 2).

Spherical harmonics are inherently better suited for spheres than DFS-based encodings that assume a rectangular domain between longitude and latitude. As shown in Figure 4b), this assumption does not hold at high latitudes, where DFS-based encoders drop systematically in accuracy, while SH remains accurate.

Reviewer: The gains from SH are only large when using a low-capacity network. [...]. To me, this seems to suggest that "expressiveness" is the main issue.

An expressive high-capacity neural network, such as FcNet or Siren, can compensate for inaccuracies of sub-optimal DFS-based embeddings. In our view, the good performance of SH even on the Linear "low-capacity network" highlights that spherical harmonic basis functions alone are inherently better suited for global-scale prediction tasks.

Reviewer: What's the difference in computational efficiency between SH + SirenNet and X + SirenNet where X is another high-performing input encoding?

The table below shows runtimes for the land-ocean dataset (Table 1b) $\pm$ from 5 runs. We reran the models on CPU on a Macbook. The SH embeddings in this used 10 Legendre Polynomials (determined by hyperparameter tuning as detailed in the appendix).

In practice, we see no disadvantage in terms of the runtime of one over another.

Reviewer: What's the best argument [...] that SH is worthwhile to use as a component in a real system compared to the alternatives?

SH basis functions are effective for all networks and perform well even for simple (i.e., Linear) neural networks (Tables 1,2), they can be faster than DFS encodings until L=20 (Figure 5) and are even up to L=100 not excessive (5.62 sec. run time for 2000 points on Macbook CPU).

Reviewer: What is the trade-off between performance and efficiency? [...] The important thing is the trade-off between performance and computational efficiency. E.g. is SH + Linear faster than SH + FCNet?

The runtime difference between Linear(SH) and FCNet(SH) (with L=15 polynomials) of 2k points on the land-ocean dataset CPU/Macbook is minor: Linear(SH) took 0.91 s, while FCNet(SH) took 0.97 s.

Hence, Linear(SH) is slightly faster than FCNet(SH) by 0.06 s due to the more complex FCNet model.

Below, we also provide the accuracy % and runtimes in seconds from Table 5 (inference of 10000 test points of the checkerboard dataset on GPU):

A larger L (and longer compute runtime) does not automatically lead to better results (i.e., here L=20 is better than L=40).

Reviewer: Isn't network capacity and degree of fit a significant confounder in the experiments? [...] How many layers does SirenNet have in this work?

We tune each PE-NN combination separately for 100 epochs. The hyperparameter space includes hidden dimensions within the networks, number of layers, learning rate, weight decay, number of frequencies S, $r_{min}$. The exact parameter range is detailed in the appendix.

Reviewer: Are FCNet and SirenNet matched in this sense?

In the hyperparameter tuning procedure, they are matched in the sense that we take the best-performing model configuration on the validation set.

Reviewer: Do they behave differently as the capacity of the networks increases?

We see no clear connection between model capacity and classification/interpolation performance. Simpler (lower-capacity) models also interpolate better towards spatially different test points.

Reviewer: What is the effect of training duration / how well the network has fit to the training data?

During training, we monitor the validation loss and perform early stopping with patience of 30 epochs. We do this for all NN-PE combinations consistently, as detailed in the revised appendix.

Misc Comments

We integrated the misc comments in the revised manuscript and appendix (blue font).

Thank you for the thoughtful responses!

It seems to be the case that SH is neither much faster nor much more accurate than the alternatives considered in the paper. (Please correct me if I've misunderstood.) This is not a problem in and of itself (since the paper has other merits), but it is a problem if the paper fails to make this clear to the reader. I would strongly encourage revisions for clarity along these lines.

Regardless, I think the paper has value and will be of interest to the community, so I will increase my rating to weak accept. Among other things, the careful separate analysis of encodings and networks is a great contribution.