acceptodds
Under review as a conference paper at ICLR 2027

Shape Dissimilarity for Product-Spherically Embedded Trajectories

Abstract

Traditional shape analysis in Euclidean space involves constructing pre-shapes by quotienting out Euclidean uniform scaling and translation. However, these operations are fundamentally undefined on spheres or product spheres. To overcome this geometric incompatibility, we introduce a novel framework for the shape analysis of product-spherically embedded trajectories that reside on or can be equivalently mapped to a product sphere. We propose the roduct-pherical hape issimilarity (PSSD) measure to rigorously quantify the shape differences between two such trajectories. Theoretically, we establish that the PSSD measure evaluates to zero if and only if two trajectories are equivalent up to spherical uniform scaling and reference-based spherical translation. Furthermore, we develop two algorithms: one extracts intrinsic spherical pre-shapes that remain embedded in the product sphere, and the other aligns the rotation of these pre-shapes. We also propose a rigorous smoothing estimator that recovers trajectories from discrete observations and guarantees uniform convergence. Extensive empirical evaluations across various synthetic and real-world datasets demonstrate that the PSSD measure achieves superior discriminative power in complex non-Euclidean classification tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Non-Euclidean data, Procrustes analysis, Shape analysis

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_p7JaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Product-Spherical Shape Dissimilarity (PSSD), a framework for comparing trajectories embedded in spheres or product spheres while accounting for spherical scaling and reference-based translation. It constructs spherical pre-shapes, extends the framework to rotational alignment through Spherical Procrustes Dissimilarity (SPD), and develops a smoothing estimator with a uniform convergence guarantee. The authors evaluate the proposed measures on synthetic and real-world datasets using k-nearest-neighbor classification and comparisons with geometric and neural-network baselines. The results show consistent gains on the designed synthetic settings, while improvements on real-world datasets are more mixed.

Critical strengths

(1) A potentially valuable geometric formulation. The paper introduces PSSD to compare trajectories on spheres and product spheres while accounting for specific scaling and reference-based alignment invariances. This provides a mathematically explicit framework for a class of non-Euclidean shape analysis problems and represents a potentially meaningful methodological contribution.

(2) Theoretical analysis beyond a standalone distance definition. The paper extends the proposed framework to rotational alignment and develops a smoothing estimator with a uniform convergence guarantee. The combination of geometric formulation, theoretical analysis, and evaluation on synthetic and real-world datasets gives the work a coherent methodological foundation.

Critical weaknesses

(1) Insufficient justification for the proposed invariance. The paper does not adequately explain why the reference-based spherical alignment and scaling invariance are appropriate for the target applications, or how the choice of reference center affects PSSD.

(2) Limited evidence of generalization. The synthetic experiments closely match the invariance that PSSD is designed to capture, while the improvements on real-world datasets are mixed. Comparisons with closely related manifold-based shape analysis methods are also missing, making it difficult to assess the practical advantage of PSSD over existing approaches.

Questions for authors

(1) How sensitive is PSSD to the choice of reference center? Can the authors explain how the reference center is estimated during evaluation and whether changing the reference set affects pairwise dissimilarities?

(2) Can the authors compare PSSD with more closely related manifold-based trajectory or shape-analysis methods, and evaluate settings where class differences are not primarily driven by the scaling transformations that PSSD is designed to remove?

2 comments

We thank the reviewer for these comments. We address the concerns below.

1. Motivation for spherical alignment and scaling invariance

The proposed invariances target applications in which the geometric pattern of a trajectory is of interest, while the location and scale are treated as nuisance. Besides the shape classification considered in the paper, data alignment is also an important application in the literature (see, e.g., Dryden and Mardia, 2016). Moreover, other real-data examples also motivate spherical shape analysis, such as flight trajectories. For regional flight segments represented by geographical positions on a sphere, one may wish to identify recurring looping or bending patterns across different locations. The proposed method enables comparison of these patterns irrespective of location and size.

2. Reference point and dissimilarity matrix calculation

The common reference is part of the definition of PSSD and is estimated as the Fréchet mean of the individual trajectory centers. The PSSD measure depends on the reference point, but when the common reference point is fixed, the zero-dissimilarity relation is transitive for different spherical trajectories; see 5. Reference point sensitivity analysis of our response to Reviewer Zq28 for details.

In our reported evaluation, the reference is estimated once from the full data collection, and the resulting dissimilarity matrix is held fixed across the classification splits. Classifier fitting and tuning use training labels only. This is a classical procedure for semi-supervised tasks in the multidimensional scaling literature, where the full dissimilarity matrix is obtained first; see, e.g., Section 3 of Trosset et al. (2008).

The sensitivity experiment in 5. Reference point sensitivity analysis of our response to Reviewer Zq28 quantifies the effect of perturbing the reference. Over the range 00–0.200.20 radians, the mean within-group dissimilarity increases from 0.000.00 to 0.020.02, while classification remains at 100100\\%. These results indicate classification stability in the simulation setting examined.

3. Relationship to manifold-based trajectory analysis

The closest related literature concerns elastic shape analysis of manifold-valued trajectories. These approaches, however, address a different topic on the temporal registration. Our method is developed in the spirit of Procrustes shape analysis. In 2. Relationship to existing elastic shape analysis methods of our response to Reviewer Zq28, we discuss these differences in detail and provide an analytical example in which the proposed measure gives zero dissimilarity while the existing measures give positive dissimilarity. This illustrates that the methods impose different notions of shape equivalence.

To the best of our knowledge, this specific Procrustes-type formulation for product-spherical trajectory shapes has not previously been developed.

In 3 Additional synthetic experiments of our response to Reviewer yWNV, we provide additional experiments on the comparison with the existing elastic shape analysis methods of manifold-valued trajectories in Su et al. (2014) and Zhang et al. (2018). In both the single- and product-sphere settings, PSSD gives zero within-group dissimilarity and positive between-group dissimilarity, with 100100\\% classification accuracy. The two elastic dissimilarities have similar average within-group and between-group values and lower classification accuracy in this controlled example. These results suggest that the two measures do not fully capture the notion of spherical shape considered in this work. The analytical comparison in 2. Relationship to existing elastic shape analysis methods of our response to Reviewer Zq28 explains this distinction.

4. Data-generation procedure

We thank the reviewer for raising this concern. Our primary objective is to quantify spherical trajectory shape under the specified equivalence with respect to spherical translation and uniform scaling. The synthetic study therefore requires known shape classes. Generating trajectories from a common underlying shape within each group, while varying the centers and scales, which are treated as nuisance, provides interpretable ground truth for this evaluation. Classification serves as a downstream assessment of whether the dissimilarity measures could recover the structure. The synthetic results should be interpreted as a controlled assessment of recovery under the specified shape equivalence, which is a common experimental design in the shape analysis literature.

References

  • Dryden, I. L. and Mardia, K. V. (2016). Statistical Shape Analysis: With Applications in R. 2nd edition. John Wiley & Sons.
  • Su, J., Kurtek, S., Klassen, E., and Srivastava, A. (2014). Statistical analysis of trajectories on Riemannian manifolds: Bird migration, hurricane tracking and video surveillance. The Annals of Applied Statistics, 8(1), 530–552.
  • Trosset, M. W., Priebe, C. E., Park, Y., and Miller, M. I. (2008). Semisupervised learning from dissimilarity data. Computational Statistics & Data Analysis, 52(10), 4643–4657.
  • Zhang, Z., Klassen, E., and Srivastava, A. (2018). Phase-amplitude separation and modeling of spherical trajectories. Journal of Computational and Graphical Statistics, 27(1), 85–97.
Reviewer_yWNVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a geometric framework for comparing the shapes of trajectories on spheres and product spheres. Its main contribution is the Product-Spherical Shape Dissimilarity (PSSD) measure, which combines Fréchet centering, reference-based spherical alignment, and global normalization to achieve invariance to spherical translation and uniform scaling. The authors establish theoretical properties of PSSD, provide algorithms for spherical pre-shape extraction and an additional rotation-alignment method, and introduce a smoothing estimator with a convergence guarantee. Experimental results on synthetic and real-world datasets show encouraging classification performance.

Critical strengths
  1. The paper addresses an important problem in analyzing non-Euclidean trajectories, where traditional Euclidean shape-analysis methods cannot be directly applied.
  2. The proposed framework is mathematically well motivated, and the authors provide theoretical results and proofs to support their main claims.
  3. PSSD consistently outperforms the baselines in the synthetic experiments and achieves promising results on several real-world datasets.
Critical weaknesses

W1. Novelty over existing approaches is not fully clear. The framework builds on established ideas, including Fréchet centering, operator-based spherical differences (Section 3.1), global normalization (Definitions 2–3), and an optional Procrustes-based rotation-alignment method (Section 4.1). While their combination is useful, the paper could more clearly explain the new mathematical contribution, particularly for product spheres. W2. The assumptions behind Theorem 4 need clarification. The proof in Appendix H uses compactness and total boundedness, but these properties do not generally hold for infinite-dimensional Hilbert spheres, which the paper also considers. It is unclear whether Assumptions A1–A4 are sufficient or whether additional conditions are needed for the convergence guarantee. W3. Synthetic evaluation of product spheres is limited. The synthetic experiments focus on single-sphere trajectories rather than general product spheres. Moreover, Algorithm 1 is used in data generation (Appendix I.5), which may favor PSSD. Experiments involving multiple spherical components and alternative data-generation procedures would better support the broader contribution. W4. Some real-world improvements are modest. Although PSSD performs well overall, its improvements over simpler baselines are small on some datasets. For example, PSSD achieves 95.64% accuracy on LTDC compared with 95.45% for SD under random splitting, while several methods achieve 100% on MultiGait. Reporting standard deviations or confidence intervals across the repeated splits would help assess the reliability of these gains.

Questions for authors

The paper addresses an interesting problem and presents a mathematically motivated framework with encouraging experimental results. I appreciate the theoretical analysis and the evaluations on both synthetic and real-world datasets. However, I have some concerns about the novelty of the proposed formulation relative to existing methods, the assumptions behind the convergence guarantee, and the limited synthetic evaluation of general product-spherical trajectories. These issues make it difficult to fully assess the strength and generality of the main contributions.

  1. What is the key mathematical advantage of PSSD over existing spherical or tangent-space shape-analysis approaches, particularly in the product-sphere setting?
  2. How is the common reference center selected, and how sensitive is PSSD to this choice? Please also clarify whether any shared preprocessing parameters are estimated from training data only.
  3. Can you clarify the compactness assumptions used in Theorem 4 and whether the convergence guarantee extends to infinite-dimensional Hilbert spheres?
2 comments

We thank the reviewer for the constructive comments. We address the concerns below.

1. Novelty and mathematical advantage

We thank the reviewer for requesting a clearer account of the contribution. Our contribution is to formulate a reference-based notion of product-spherical trajectory shape and construct a dissimilarity with an explicit characterization of its zero-dissimilarity equivalence. The underlying geometric identities and theoretical properties build on established mathematical results.

The closest related literature concerns elastic shape analysis of manifold-valued trajectories. These approaches, however, address a different topic: they use velocity-based representations for temporal registration and phase-amplitude analysis, whereas our method is developed in the spirit of Procrustes shape analysis. In 2. Relationship to existing elastic shape analysis methods of our response to Reviewer Zq28, we discuss these differences in detail and provide an analytical example in which the proposed measure gives zero dissimilarity while the existing measures give positive dissimilarity. This illustrates that the methods impose different notions of shape equivalence.

To the best of our knowledge, this specific Procrustes-type formulation for product-spherical trajectory shapes has not previously been developed.

2. Validity of Theorem 4

Theorem 4 covers both finite- and infinite-dimensional spheres. The proof arguments involving compactness are built on the recent theoretical developments in Theorems 1 and 2 of Xu et al. (2025). That is, for uniform consistency over an infinite-dimensional sphere, it suffices to show uniform consistency over any compact subset of the infinite-dimensional sphere. Such localization arguments are also standard in classical empirical process theory; see van der Vaart and Wellner (1996, Corollary 3.2.3) and Kosorok (2008, Lemma 14.3).

3. Additional synthetic experiments

We conduct additional synthetic experiments on the sphere S2\mathcal{S}^2 and the product sphere S2×S2\mathcal{S}^2\times\mathcal{S}^2 as requested, including comparisons with the existing elastic shape analysis methods for manifold-valued trajectories.

We generate one sample of 200 trajectories, with 100 trajectories in each group. Let μ0=(0,0,1)⊤\mu_0=(0,0,1)^\top, K=10K=10, cK=0.1/Kc_K=0.1/K, b1=0.60b_1=0.60, and b2=0.70b_2=0.70. For group g∈1,2g\in\\{1,2\\}, let wg,1(u)=cK(sin⁡(2πKu),,bgsin⁡(2πu),,0)⊤w_{g,1}(u)=c_K(\sin(2\pi Ku),\\,b_g\sin(2\pi u),\\,0)^\top and wg,2(u)=cK(0.8sin⁡(4πKu),,0.5bgsin⁡(4πu),,0)⊤w_{g,2}(u)=c_K( 0.8\sin(4\pi Ku),\\,0.5b_g\sin(4\pi u),\\,0)^\top.

For ℓ∈1,2\ell\in\\{1,2\\} and j=1,2,…,100j=1,2,\ldots,100, we generate trajectory centers as η2j−1,ℓ=Exp⁡μ0(rℓjvℓj)\eta_{2j-1,\ell}=\operatorname{Exp}_{\mu_0}(r_{\ell j}v_{\ell j}) and η2j,ℓ=Exp⁡μ0(−rℓjvℓj)\eta_{2j,\ell} =\operatorname{Exp}_{\mu_0}(-r_{\ell j}v_{\ell j}) where vℓjv_{\ell j} is uniformly distributed on the unit circle in Tμ0S2T_{\mu_0}\mathcal{S}^2 and cos⁡(rℓj)∼Unif⁡(cos⁡(0.40),1)\cos(r_{\ell j})\sim\operatorname{Unif}(\cos(0.40),1). We generate the scales aia_i as log⁡ai∼Unif⁡(log⁡(0.05),log⁡(2))\log a_i\sim\operatorname{Unif}(\log(0.05),\log(2)) for i=1,2,…,200i=1,2,\ldots,200. The balanced group labels gi∈1,2g_i\in\\{1,2\\} are randomly assigned independently of the centers and scales. Let RiℓR_{i\ell} be the rotation matrix taking ηiℓ\eta_{i\ell} to μ0\mu_0. The spherical trajectories are generated via

fiℓ(u)=Riℓ⊤Exp⁡μ0aiwgi,ℓ(u),ℓ∈1,2.f_{i\ell}(u)=R_{i\ell}^{\top}\operatorname{Exp}_{\mu_0}\\{a_iw_{g_i,\ell}(u)\\},\qquad \ell\in\\{1,2\\}.

For the single-sphere setting, we use fi1f_{i1}, whereas the product-sphere setting uses (fi1,fi2)(f_{i1},f_{i2}).

This design ensures that the two groups have different spherical shapes while the within-group trajectories have the same spherical shape. We now compare our method with the existing elastic shape analysis methods in Su et al. (2014) and Zhang et al. (2018).

The table below reports the average within- and between-group dissimilarities, together with the classification accuracy using the dissimilarity matrix for the three measures. As expected, PSSD gives a zero value for within-group trajectories and results in a positive value for the between-group ones. This leads to 100% classification accuracy. However, for both elastic methods, the within-group and between-group dissimilarity measures return similar values, and classification accuracy is lower.

Table 3. Average dissimilarity measures and classification accuracy (%) for one sample.

SettingMethodWithin-groupBetween-groupRandom split (%)LOOCV (%)
Single spherePSSD0.000.07100.00100.00
Su et al. (2014)0.240.2460.0070.50
Zhang et al. (2018)0.450.4552.0042.00
Product spherePSSD0.000.07100.00100.00
Su et al. (2014)0.340.3460.0060.00
Zhang et al. (2018)0.660.6634.0048.00

4. Data-generation procedure

We thank the reviewer for raising this concern. Our primary objective is to quantify spherical trajectory shape under the specified equivalence with respect to spherical translation and uniform scaling. The synthetic study therefore requires known shape classes. Generating trajectories from a common underlying shape within each group, while varying the location and scale, which are treated as nuisance, provides interpretable ground truth for this evaluation. Classification serves as a downstream assessment of whether the dissimilarity measures could recover this structure. The synthetic results should be interpreted as a controlled assessment of recovery under the specified shape equivalence, which is a common experimental design in the shape analysis literature.

5. Real data performance analysis

We agree that the improvements over simpler baselines are modest in some real-data comparisons. Our primary objective is to develop a dissimilarity measure for spherical trajectory shapes, with classification serving as an assessment of its practical usefulness. Although the baseline measures do not give zero dissimilarity for trajectories with the same spherical shape, it may not reduce the classification performance significantly. In shape analysis, differences in the underlying pre-shapes can also lead to substantial differences in the raw configurations. This can create a large gap between within-group and between-group dissimilarities, allowing classifiers to maintain high accuracy even when the measures retain variation that the intended shape definition treats as nuisance.

6. Reference point and dissimilarity matrix calculation

For trajectories fS,ii=1n\\{f_{{\mathrm{S}},i}\\}_{i=1}^n, the common reference is estimated as

μS=arg,min⁡μ∈Sn−1∑i=1ndS2(μ,fˉS,i),\mu_{\mathrm{S}}=\operatorname*{arg\\,min}_{\mu\in\mathcal S}n^{-1}\sum_{i=1}^nd_{\mathcal S}^{2}(\mu,\bar f_{{\mathrm{S}},i}),

where fˉS,i\bar f_{{\mathrm{S}},i} is the Fréchet center of the iith trajectory.

In the reported dissimilarity-matrix evaluation, this reference point is estimated once from the full unlabeled collection. The reference and resulting dissimilarity matrix are then held fixed across random splits and cross-validation folds. Classifier training and tuning use training labels only. This is a classical procedure for semi-supervised tasks in the multidimensional scaling literature, where the full dissimilarity matrix is obtained first; see, e.g., Section 3 of Trosset et al. (2008).

The dissimilarity generally depends on the reference point. In the sensitivity experiment reported in 5. Reference point sensitivity analysis of our response to Reviewer Zq28, classification accuracy remains at 100100\\% for reference perturbations up to 0.200.20 radians, although the dissimilarities change. This suggests that classification is stable under moderate perturbations of the reference point.

References

  • Kosorok, M. R. (2008). Introduction to Empirical Processes and Semiparametric Inference. Springer, New York.
  • Su, J., Kurtek, S., Klassen, E., and Srivastava, A. (2014). Statistical analysis of trajectories on Riemannian manifolds: Bird migration, hurricane tracking and video surveillance. The Annals of Applied Statistics, 8(1), 530–552.
  • Trosset, M. W., Priebe, C. E., Park, Y., and Miller, M. I. (2008). Semisupervised learning from dissimilarity data. Computational Statistics & Data Analysis, 52(10), 4643–4657.
  • van der Vaart, A. W., and Wellner, J. A. (1996). Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York.
  • Xu, J., Wood, A. T. A., and Zou, T. (2025). Robust functional principal component analysis for non-Euclidean random objects. Biometrika, 112(4), asaf040.
  • Zhang, Z., Klassen, E., and Srivastava, A. (2018). Phase-amplitude separation and modeling of spherical trajectories. Journal of Computational and Graphical Statistics, 27(1), 85–97.
Reviewer_Zq28OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a method to quanitfy diferences between shapes of trajectories. The approach is to represent the data of shapes either on a sphere or a product of spheres. They also propose a measure, the Product-Spherical Shape Dissimilarity (PSSD) measure to quantify the shape differences between trajectories. They show that the PSSD measure yields a zero distance iff two trajectories are equivalent up to spherical uniform scaling and translation. They also describe algorithms to extract intrinsic spherical pre-shapes that on the product sphere, and to rotationally align the pre-shapes in addition to a method for performing smoothing of the trajectories. They show experimental results on both synthetic and real datasets of compositional data, age distributional time series and biomechanial data from human gait.

Critical strengths

The paper tackles an important problem of comparing shapes of time-series data.

The formulation is well defined and the paper is clearly written.

The theoretical results are sound.

The method is demonstrated on both synthetic and real datasets and the authors have released their software.

Critical weaknesses

One of the main weaknesses is lack of adequate motivation through experiments for representing data on product of spheres. The idea of spherical shape space is sound and experimental results exist to justify that. However, the product of spheres idea, although mathematically sound, is not adequately justified. The gait angle data is meant to be represented on the product of spheres, however it is not shown whether the product of spheres representation is needed here.

A few closely related methods are not cited. Instead different works from the same groups of authors are cited. For e.g. The authors cite Joshi, Jingyong Su, Zhengwu Zhang et al. Elastic shape analysis of functions, curves and trajectories. However, they should also cite those authors main papers including the ttransported squre root velocity functions (TSRVF) Su et al. (2014), Zhang et al. (2018, arXiv:1603.07066), etc. In those works, the authors covered reparameterization invariance. The authors should discuss these works and point out hte similarities and differences.

The authors claim that "The established elastic shape framework was primarily developed for Euclidean trajectories, utilizing pre-shape spaces to ensure invariance under uniform scaling and translation (Srivastava et al., 2010; Srivastava & Klassen, 2016). However, such operations are fundamentally ill-defined on product spheres, which lack a global Euclidean vector space structure. " However, those representations can be extended to product spaces and one can define a Riemannian metric on the product of those spaces. It is true that these product spheres if constructed will lack a global Euclidean vector space structure. But that is true in this method as well. For e.g. in the proposed work, the authors are still exploiting the tangent space structure similar to the above methods in definiining baseline spherical shapes. Additionally, the claim that translation and scaling are "fundamentally ill-defined" on spheres is not true. The SRVF representation is invariant to translation and scale and so are the TSRVF representations.

The proposed PSSD measure depends on a reference point used to align the trajectories (here it's the mean). This mean is estimatd from a population, and thus the notion of this dissimilarity measure changes if the population changes. This is a limitation since this formulation turns the PSSD and SSD into a population based measure.

The theoremss 1 and 2 follow from established results and not strictly the main contributions of the paper. For e.g. Theorem 1 follows from Cauchy-Schwarz inequality (as the authors state) and spherical scaling (norming). Theorem 3 directly follows from the procrustes solution. The derivation for optimal transport in Appendix B directly follows from parallel transport and Rodrigues formula and that the idea of exponential maps to project the tangent vectors in a tangent space back to the sphere has a closed form on a sphere. These results can be included for completeness, but are not necessarily the main contributions of the paper.

In the experimental results, the product sphere formulation is not well justified. Here, on the multigait dataset, several simpler Euclidean and spherical dissimilarity measures achieve superior classification accuracy. Thus this experiment or the dataaset that is chosen doesn't show a clear advantage for the product of spheres approach. This leads to another question about the product metric. Currently the authors use equally weighted metrics to construct the full product metric. While this is fine, generallly the application may require unequal weights attached to the product metric. For e.g. in the gait case, some articulations (joint angles) may be need to be penalized more compared to others. Likewise, for another application, one may deal with another weighted metric. While this doesn't change the theoretical formulations, it may change experimental results.

Again, in the experimental results, the SPDs and PSSD perform the best on synthetic data, however, are only marginally better than the other methods including Euclidean shape similairty on the real datasets. Further, the smoothed results seem to be substantially better than the unsmoothed results.

Questions for authors

It is clear that PSSD could change with the Frechet mean. How are the Frechet means computed in practice? Are they computed iteratively and converge to a local solution? Have you conducted any sensitivity analysis as to how the metric changes with perturbations in the mean?

It will be useful to discuss the papers by Su et al. 2014 Ananls of Applied Stats, and Zhang et al. 2018 (Phase-Amplitude Separation and Modeling of Spherical Trajectories).

Similarly, for experimental results, it would be useful to discuss as to how sensitive the classification results are to the choice of reference point. Is this reference point estimated from the training data and kept fixed or is it repeatedly estimated during cross-validation?

In the experimental results, the SPDs and PSSD perform the best on synthetic data, however, are only marginally better than the other methods including Euclidean shape similairty on the real datasets. Further, the smoothed results seem to be substantially better than the unsmoothed results. Can the authors explain?

5 comments

We thank the reviewer for the constructive comments and for recognizing the importance of the problem, the clarity of the formulation, and the soundness of the theoretical results. We address the concerns below.

1. Motivation for the product-spherical formulation

The product-spherical extension was initially motivated by human gait data, where each trajectory comprises several distinct joint-angle components. Representing each angular component on a circle naturally gives a trajectory on a product of circles. MultiGait also provides experimentally defined ground-truth class labels corresponding to normal walking and walking with an immobilized knee or ankle.

Our purpose in including this dataset was to assess whether the proposed shape dissimilarity retains sufficient information to distinguish these known biomechanical conditions. In particular, because standard methods already achieve perfect classification on this dataset, it provides a useful empirical check that the proposed geometric normalization also preserves the discriminatory information needed for this task. PSSD achieves perfect classification in the reported evaluations, matching the strongest baselines.

Other applications could also motivate accommodating multiple spherical components. For example, the unit directions of multiple bones in three-dimensional skeletal motion form trajectories on a product of two-dimensional spheres. Similarly, several jointly observed compositional trajectories admit a product-spherical representation after applying a square-root transformation separately to each composition.

We will add these examples and clarify that MultiGait evaluates the practical usefulness of the proposed method, while the broader motivation for the product formulation comes from data comprising multiple components with their own spherical constraints. Moreover, we would like to clarify that classification based on shape dissimilarity measures that we study in the numerical experiments is just one application of shape analysis. Other important applications include data alignment (see, e.g., Dryden and Mardia, 2016).

2. Relationship to existing elastic shape analysis methods

We thank the reviewer for highlighting Su et al. (2014) and Zhang et al. (2018). We will cite these papers directly and revise the statement that translation and scaling are “fundamentally ill-defined” on product spheres. We will clarify the connections and differences between Procrustes and elastic shape analysis in the Introduction.

When temporal alignment is omitted, the classical Procrustes and elastic shape techniques in Euclidean spaces will yield the same zero-distance equivalence. This is because the square-root velocity function (SRVF) representation, after normalization, is invariant to translation and scale. However, this property does not generally hold in the sphere setting. To support this statement, in the following, we first revisit the connection between the classical Procrustes shape analysis and elastic shape analysis, then summarize the elastic shape analysis methods of Su et al. (2014) and Zhang et al. (2018), and finally discuss the difference between our method and the ones in Su et al. (2014) and Zhang et al. (2018).

Classical Procrustes versus elastic shape analysis. Classical Procrustes shape analysis in Euclidean spaces compares positional configurations with specified point correspondences, removing translation, uniform scale, and rotation (Kendall, 1984; Goodall, 1991). For curves f,g:[0,1]→Rmf,g:[0,1]\to\mathbb{R}^m, the dissimilarity measure dPd_{\mathrm P} is constructed to satisfy:

\quad\text{for all }u,$$ for some $a > 0$, $R\in\operatorname{SO}(m)$, and $b\in\mathbb{R}^m$. Here, $\operatorname{SO}(m)=\\{R\in\mathbb{R}^{m\times m}:R^\top R=I,\\ \det R=1\\}$. Elastic shape analysis in Euclidean spaces compares parameterized curves using a differential metric and optimizes their point correspondences through temporal reparameterization, while also removing translation, uniform scale, and rotation in the Euclidean setting (Srivastava et al., 2010; Needham and Kurtek, 2020). For regular curves $f,g:[0,1]\to\mathbb{R}^m$, when the optimal alignment is attained by an admissible reparameterization, the elastic shape dissimilarity $d_{\mathrm{EL}}$ satisfies $$d_{\mathrm{EL}}(f,g)=0\quad\Longleftrightarrow\quad f(u)=aR g(\gamma(u))+b \quad\text{for all }u,$$ for some $a > 0$, $R\in\operatorname{SO}(m)$, $b\in\mathbb{R}^m$, and $\gamma\in\Gamma$. Here, $$\Gamma=\left\\{\gamma:[0,1]\to[0,1]:\gamma\text{ is a }C^1\text{ diffeomorphism}, \\ \gamma(0)=0,\\ \gamma(1)=1,\\ \dot\gamma > 0\right\\},$$ and $\dot\gamma(u)=d\gamma(u)/du$ denotes the derivative of the time-warping function $\gamma$. Thus, elastic shape analysis extends the comparison to accommodate temporal reparameterization. When $\gamma(u)=u$ is fixed and the same spatial invariances are imposed, classical Procrustes and elastic shape analysis identify the same Euclidean similarity classes.

2. Relationship to existing elastic shape analysis methods (continued)

Elastic shape analysis for manifold-valued trajectories. Now we consider f,g:[0,1]→S2f,g:[0,1]\to\mathcal{S}^2 and let TxS2=v∈R3:x⊤v=0T_x\mathcal{S}^2=\\{v\in\mathbb{R}^3:x^\top v=0\\}. For a function ff, define f˙(u)=df(u)/du∈Tf(u)S2\dot f(u)=df(u)/du\in T_{f(u)}\mathcal{S}^2 as the velocity vector of ff at time uu. Following the comparison above, we focus on the case where temporal alignment is unnecessary. That is, we fix the identity reparameterization γ(u)=u\gamma(u)=u and omit optimization over time warpings.

Su et al. (2014) develop temporal registration and phase-invariant statistical analysis for manifold-valued trajectories. Specializing their representation to S2\mathcal{S}^2, their transported square-root vector field (TSRVF) is

\right\\}\in T_c\mathcal{S}^2.$$ Here, $c\in\mathcal{S}^2$ is a common fixed reference and $\mathcal P^{\mathrm{geo}}_{f(u)\to c}$ denotes parallel transport along the unique shortest connecting geodesic. Without temporal alignment, their dissimilarity $d_{\mathrm{Su}}$ is $$d_{\mathrm{Su}}(f,g)=\left\\{\int_0^1\\|h_f(u)-h_g(u)\\|^2\\,du\right\\}^{1/2},$$ which satisfies $$d_{\mathrm{Su}}(f,g)=0\quad\Longleftrightarrow\quad h_f(u)=h_g(u) \quad\text{for all }u.$$ Thus, zero dissimilarity requires identical transported velocity representations at corresponding times. Notably, when $f(0)=g(0)$, equality of the TSRVFs implies $f(u)=g(u)$ for all $u$. Zhang et al. (2018) develop phase-amplitude separation and intrinsic statistical analysis of spherical trajectories. For a function $f$, let $p_f=f(0)$ and $q_f(u)=\mathcal P^{f}_{f(u)\to p_f} \left\\{\dot f(u)/\sqrt{\\|\dot f(u)\\|}\right\\}\in T_{p_f}\mathcal{S}^2$, where $\mathcal P^{f}_{f(u)\to p_f}$ denotes parallel transport backwards along the trajectory $f$. Without temporal alignment, their dissimilarity $d_{\mathrm{Zhang}}$ satisfies $$\begin{aligned} d_{\mathrm{Zhang}}(f,g)=0 &\quad\Longleftrightarrow\quad p_f=p_g \\ \text{and}\\ q_f(u)=q_g(u)\\ \text{for all }u\\\\ &\quad\Longleftrightarrow\quad f(u)=g(u)\\ \text{for all }u. \end{aligned}$$ Therefore, the measure in Zhang et al. (2018) is zero only when the spherical trajectories are equal at every point. *Difference between the PSSD measure and the existing measures.* The distinction between the PSSD measure and the existing measures in Su et al. (2014) and Zhang et al. (2018) can be demonstrated using two trajectories on a single sphere. Let $e_1=(1,0,0)^\top$, $e_2=(0,1,0)^\top$, and $\mu=(0,0,1)^\top$. Choose $0 < b,r < \pi/8$ and define $\nu_-=-\sin b\\,e_2+\cos b\\,\mu$, $\nu_+=\sin b\\,e_2+\cos b\\,\mu$, and $s(u)=r(2u-1)$. Consider two spherical trajectories of the same spherical shape $f(u)=\cos s(u)\\,\nu_-+\sin s(u)\\,e_1$ and $g(u)=\cos s(u)\\,\nu_++\sin s(u)\\,e_1$. Their velocities are $\dot f(u)=2r\\{-\sin s(u)\\,\nu_-+\cos s(u)\\,e_1\\}$ and $\dot g(u)=2r\\{-\sin s(u)\\,\nu_++\cos s(u)\\,e_1\\}$, respectively. For the proposed PSSD measure, one can verify that $D_{\mathrm{S}}(f,g)=0$. However, calculations show that $d_{\mathrm{Su}}(f,g) > 0$ and $d_{\mathrm{Zhang}}(f,g) > 0$. This example demonstrates that when a change of spherical location occurs, even though the two trajectories have the same shape, the existing distances in Su et al. (2014) and Zhang et al. (2018) fail to give zero dissimilarity values. In **3. Additional synthetic experiments** of our response to Reviewer yWNV, we provide additional experiments on the comparison with these two measures, where the temporal alignment step is applied as in Su et al. (2014) and Zhang et al. (2018). These results support the statement that the two measures do not fully capture the notion of spherical shape considered in this work. **3. Theoretical contributions** We agree that Theorems 1–3 and the geometric identities in Appendix B are established on known mathematical results. Their role is to support the proposed methodology by characterizing its invariance and zero-dissimilarity properties and establishing the stated rotation-alignment solution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We are encouraged by their recognition of the key contributions and strengths of our work.

We have carefully addressed each individual comment provided by the reviewers and believe we have successfully resolved most of their concerns. During the rebuttal phase, we expanded our discussions, incorporated the suggested experiments, and made relevant updates to further strengthen our manuscript. Below, we summarize the core contributions of our study, alongside in-depth discussions during the rebuttal, and the experimental additions.

Core contributions

  • Methodology. We introduce a Procrustes framework for comparing the shapes of trajectories on products of spheres. The framework constructs intrinsic pre-shapes that remain on the product sphere and provides a smoothing estimator for trajectories observed at discrete times.
  • Theory. We characterize when the product-spherical shape dissimilarity (PSSD) is zero. Two trajectories have zero dissimilarity if and only if they are equivalent up to spherical translation, uniform scaling, and optional rotation. We also establish uniform convergence of the smoothing estimator under the temporal dependence.

Updates of in-depth discussions in the revised manuscript

  • Relationship to elastic shape analysis (R1, R2, R3). The Introduction and Section 3.1 explain the difference between our formulation and the existing manifold elastic methods. These elastic methods align the timing along a path to account for differences in traversal speed. Appendix B gives details and an analytical example in which two trajectories at different spherical locations have zero PSSD but positive elastic dissimilarities.
  • Motivation and choice of MultiGait (R1, R2, R3). We explain why we need to consider trajectories on a product of circles. MultiGait provides experimentally defined labels that let us check whether the proposed shape normalization preserves the information needed to distinguish known gait conditions.
  • Classification procedure (R1, R2, R3). Appendix J.2 clarifies that we construct the dissimilarity matrix from the full collection of trajectories without using the class labels and hold it fixed across splits. We use the training submatrix to select the number of neighbors by leave-one-out cross-validation and the test-to-training entries to predict held-out labels. This is analogous to the classical semi-supervised multidimensional scaling based on a precomputed dissimilarity matrix.

Additional experiments

  • Comparison with manifold elastic methods (R1, R2, R3). Appendix B reports comparisons with manifold elastic methods. In these simulations, PSSD gives zero dissimilarity within each shape class and positive dissimilarity between classes while both elastic measures fail to identify the same spherical shapes.
  • Product-sphere evaluation (R2). Appendix B extends the comparison from a single sphere to a product of two spheres.
  • Sensitivity to the Fréchet mean and common reference (R1). Appendices J.8 and J.9 conduct sensitivity analysis for the Fréchet mean and the common reference in the dissimilarity measure.

We believe these additions and clarifications comprehensively address the reviewers' concerns and enhance the overall quality of our manuscript.

We look forward to the reviewers' favourable consideration and remain grateful for their valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.