acceptodds
Under review as a conference paper at ICLR 2027

Comparing dynamical operators with flags

Abstract

Many scientific and engineering domains, including neuroscience, fluid dynamics, mechanics, and climate science, study ensembles of multidimensional spatiotemporal fields through spectral or modal decompositions. A recurring challenge is to compare such decompositions across many realizations while respecting both their intrinsic geometric characteristics and their hierarchy of dynamical timescales. Euclidean distances are common across these disciplines, but they are sensitive to arbitrary basis choices, whereas representing all modes by a single subspace discards their timescale organization. We represent each realization by a flag: a nested sequence of subspaces formed by grouping dynamical modes by spectral conjugacy and ordering them by decay time. We equip the space of flags with a chordal distance that lets us compare decompositions across realizations. The resulting distance is gauge-invariant by construction; we prove that it defines a metric, admits an isometric Euclidean embedding (thus enabling standard kernel methods), and is stable under perturbations of the fitted operator away from decay-time crossings. We illustrate the approach in climate science on 216 realizations from four CMIP6 models and a reanalysis across three variables, as well as synthetic ensembles with known ground truth. The flag geometry separates dynamical structures that the Euclidean baseline fails to distinguish and regularly outperforms a flat-subspace representation

Primary area:
learning on time series and dynamical systems
Keywords:
flag manifolds, dynamic mode decomposition, gauge-invariant subspace distance, kernel PCA, climate model ensembles, Koopman

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gBzeOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a distance metric for comparing dynamical operators by constructing persistence-ordered invariant flags, or nested sequences of invariant subspaces ordered by decay time. By leveraging chordal distances and establishing an isometric Euclidean embedding, the proposed method provides a gauge-invariant metric capable of comparing operators with differing subspace dimensions. The overall idea and its theoretical foundation are highly elegant.

Critical strengths
  1. Principled Geometric Framework: The proposed approach elegantly resolves the trade-off between eliminating gauge ambiguity and preserving timescale hierarchies (decay times) via persistence-ordered invariant flags (Propositions 1 and 2).

  2. Clear Domain Motivation: Combining the method with DMDc effectively isolates external forcing from internal dynamics, providing a compelling motivation for climate data analysis.

  3. Informative Visualization: Figure 1 offers an exceptionally clear and intuitive schematic overview of the proposed framework.

Critical weaknesses
  1. Over-reliance on Indirect Clustering Metrics: The empirical validation relies heavily on the Silhouette Score, which measures cluster separability rather than whether the intrinsic global geometry or continuous distance relationships are faithfully preserved.

  2. Sensitivity to Observation Noise: In the synthetic experiments (Figure 2, bottom), the Flag distance yields a lower Adjusted Rand Index (ARI) than the naive Raw A baseline under increasing observation noise, revealing potential algebraic sensitivity during Schur decomposition and eigenvalue sorting.

  3. Insufficient Validation on Real Climate Data: In Section 3.2 (CMIP6), the evaluation remains limited to cluster separability via Silhouette Scores, leaving it unproven whether the proposed metric captures meaningful climate-physical similarities.

Questions for authors
  1. Direct Metric Space Validation: Since Proposition 1 establishes an isometric Euclidean embedding, could you evaluate the global metric accuracy directly such as by examining the eigenvalue spectrum (explained variance ratio) of the Gram matrix derived via Young-Householder transformation (cMDS), rather than relying solely on Silhouette Scores?

  2. Robustness Under Observation Noise: What explains the lower Adjusted Rand Index (ARI) of the Flag distance compared to Raw A under observation noise in Figure 2? Does this stem from algebraic sensitivities during Schur decomposition or eigenvalue sorting?

  3. Inappropriateness of the Euclidean Baseline: Why does the naive Euclidean baseline yield a score close to zero across almost all panels in Figure 2? Given its consistently poor performance, is it a suitable baseline for evaluating dynamical operator comparisons?

  4. Physical Grounding in Climate Data: Beyond cluster separability in Section 3.2, is there qualitative or quantitative evidence demonstrating that the Flag distance preserves meaningful climate-physical relationships across CMIP6 models and reanalysis data?

3 comments

Thank you for calling the idea and its theory elegant, and for highlighting how flags reconcile gauge invariance with the timescale hierarchy, the DMDc motivation, and Figure 1.

W1 and Q1. We have added the suggested classical MDS analysis (App. J, Fig. 13). Because dFLd_{FL} is a Frobenius distance between tuples of projectors, cMDS on all positive eigenvalues reproduces every pairwise distance exactly. On CMIP6, the negative mass is zero to machine precision, which confirms the Euclidean embedding of Prop. 1 numerically. The spectrum also reflects cluster structure: four ESM centroids span at most three coordinates, and there the flag carries more variance than the Grassmannian on every variable (36% against about 30% on OSAT). This ratio only bounds the variance due to clusters from above, so we also report the exact between/within partition from the same Gram matrices. Model identity explains 35% of the flag variance on OSAT and SAT (Grassmannian: 29% and 28%), while the Euclidean baseline stays at the chance level of 1%. On synthetic data, cluster membership explains 99% of the flag variance. Beyond silhouette, we now also report ARI on CMIP6 (Fig. 4b).

W2 and Q2. We now run the observation-noise ablation at a shorter trajectory length T=50T=50 (Fig. 2). This is a harder setting with operators less fit on the dataset. Yet, in this setting, Flag keeps the highest silhouette and matches or exceeds every baseline in ARI. We do not hide the fact that Flag is worse than Raw A for longer trajectories. New Fig. 5 (App. G) shows that the ARI gap depends on record length: the flag leads or ties for T≤50T \leq 50, and from T=70T = 70 and above, Raw A leads at intermediate noise. The Schur step is backward stable, so this gap comes from the fitted AA itself. We hypothesize that attenuation pushes the less persistent modes toward the noise floor, so the deeper levels of the flag select different modes across trajectories. We have not yet been able to isolate the cause of the gap and report it as open in App. G.

W3 and Q4. Sec. 3.2 now interprets the distances physically. For OSAT and SAT, members of one ESM are closer to each other than to other models or to 20CRv3 (Fig. 8), so the persistent modes and their order are a structural property of each model. For SLP, the fast and weakly persistent atmospheric variability blurs these differences without erasing them. Where the distance behaves unexpectedly, the cause is dynamical, since on SAT the fast modes of CESM2 decay within about one fit step and change order across members (Fig. 12), which a single member cannot reveal. Members of one model also differ in their distance to 20CRv3 about as much as the models' medians do (Fig. 7), so no model is systematically closer to the reanalysis. Finally, the flags carry physical content. The 20CRv3 OSAT flag places two slow real modes, combining Atlantic and Pacific multidecadal variability, before an ENSO-like pair, whereas all ESM medoids and most members begin with that pair (Figs. 9 and 10). This is consistent with the underestimated multidecadal variability in models (Kravtsov et al., 2024) and with the single-member analysis of Mankovich et al. (2026), and a single subspace, which ignores the order of the modes, would miss it.

Q3. This near-zero score of the Euclidean baseline is expected. Because the clusters share eigenvalues, the only signal lies in the eigenvectors, and a coordinate-wise distance cannot align their arbitrary sign and phase. The variance analysis in App. J confirms this: cluster membership explains 3%3\% of the Euclidean variance (the chance level), compared with 99%99\% for the flag distance. We therefore keep the Euclidean baseline as a control for gauge dependence and have added SGOT as a stronger baseline. We are happy to move it to the appendix if the reviewer prefers.

I appreciate the authors’ sincere responses to my questions and the deeper analysis they carried out. While Silhouette score alone remains an indirect proxy for metric quality, the newly added PERMANOVA-style variance partitioning (R2R^2 in App. J), ARI on CMIP6 (Fig. 4b), and physical interpretations (Sec. 3.2) effectively complement it and sufficiently address my concerns. Accordingly I have upgraded my score (2->3).

We are glad to see the positive assessment of the reviewer. We remain available as we enter the last hours of the discussion period for any additional remarks.

Reviewer_R6icOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a new structure, "flag," for describing a dynamical system. A flag is an ordered nested sequence of invariant subspaces. The order of the subspaces is determined by decreasing persistence. Subsequently, this paper defines the distance between flags, proves that it can be embedded into a Euclidean space, and discusses its stability. Based on this, the author presents a method for constructing a flag using Schur decomposition. Experiments demonstrate that it exhibits better clustering capability than the baseline on both synthetic and meteorological data.

Critical strengths
  1. The motivation for this manuscript is clear, and the methodology is well-designed. The paper points out the instability of methods based on eigenvectors and addresses this issue by using nested invariant subspaces.
  2. The theoretical derivation and algorithm design are very well-developed. It proves the stability of the flag and its connection to Euclidean space. In the algorithm implementation, numerical precision and the handling of the reduced space are taken into account.
  3. This paper honestly points out the limitations of the analysis and methods. For instance, as mentioned in line 306, the theory concerning finite precision has not yet been established.
Critical weaknesses
  1. The motivations behind some of the experimental setups require further explanation. For example, in synthetic validation, why do centers share eigenvalues, and why use L=3L=3?
  2. The baselines used for comparison are rather simple; the proposed method should be compared against more advanced approaches. For instance, the paper itself mentions the SGOT method, yet it is not included in the experiments.
Questions for authors
  1. It seems that the flag only considers the ordinal relationship of the eigenvalues, disregarding their absolute magnitudes; could this lead to a decline in performance under certain conditions?
  2. In line 385/386, it claims “sign indeterminacy in the eigenvectors prevents meaningful alignment”. This requires further explanation. Why not eliminate the sign discrepancy first?
1 comment

Thank you for finding the motivation clear, the methodology and theory well developed, and for valuing our honesty about limitations.

W1. Sharing eigenvalues is deliberate: the clusters can then differ only through their invariant subspaces, which the flag compares. We test the complementary cases separately (Figs. 3 and 12). L=3L=3 is the depth of the true synthetic flag above the nuisance modes. On CMIP6, the first three levels of the 20CRv3 OSAT flag hold its slow modes and an ENSO-like pair (Apps. G and H).

W2. We added SGOT as a baseline. The flag has a higher silhouette and ARI than SGOT on every climate variable (Fig. 4b), and SGOT does not detect a crossing of decay times (Fig. 3).

Q1. Yes, we now state this tradeoff. Operators with the same invariant subspaces and ordering but different decay rates are at distance zero, which also makes dFLd_{FL} invariant to A↦cA\mathbf{A}\mapsto c\mathbf{A}. Performance can therefore drop when systems differ only in their rates. We now state this trade-off in Remark 2, and Sec. 4 notes that a spectral term can be added when rates matter.

Q2. Our method removes the sign ambiguity. A complex eigenvector is defined only up to a factor α∈C∗\alpha\in\mathbb{C}^\ast. Thus, Euclidean distances between them are sensitive to this scaling. Comparing subspaces through their projectors VV⊤V V^\top removes this freedom; therefore, we use the Euclidean baseline as a control without this step. We clarified the sentence (Sec. 3.1, App. B).

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

This submission proposes a framework for comparing trajectories of dynamical systems. The proposed method estimates a linear operator from each realization and constructs a flag—a sequence of nested invariant subspaces obtained through modal decomposition. These flags are then used to compare and cluster realizations. The claimed advantages are invariance to the choice of basis within each subspace and to eigenvector signs and phases, together with preservation of the hierarchy of modal persistence through ordering by eigenvalue magnitude. The experiments present comparisons using synthetic data and applications to data from climate models and reanalysis.

Critical strengths
  1. To my knowledge, the proposed framework is novel within this literature, and the approach seems reasonable. The argument that existing Grassmann-based methods lose the persistence ordering by combining modes into a single subspace, and that a flag representation can address this limitation, is convincing.
Critical weaknesses
  1. The experimental setup and results in Section 3.2 do not seem to sufficiently motivate the importance of the proposed method. The experiments examine whether trajectories from four climate models form groups corresponding to their source models. However, these source models are already known, and it is unclear how recovering these groupings contributes to climate research beyond validating the proposed method. The submission also measures distances between climate-model realizations and a reanalysis trajectory, but it is unclear what questions these comparisons are intended to address. Furthermore, the experimental results focus only on clustering and distances to reanalysis, without explaining what distinctive findings or interpretations the proposed method provides.

  2. The comparison with prior research and the positioning within the related literature seem insufficient. In particular, the proposed method is not experimentally compared with SGOT (Germain et al., 2026). Although I am not familiar with SGOT, the submission itself states that SGOT preserves hierarchy, gauge invariance, and basis independence. I think it could be directly relevant prior work, and the differences between the methods described in Section 3 do not seem to justify omitting an experimental comparison. Germain et al. (2026) also discuss related kernel-based methods using spectral decomposition to compare dynamical systems (Vishwanathan et al., 2007; Fujii et al., 2017; Ishikawa et al., 2018). In an application domain, Shiraishi et al. (2020) appears to use a method similar to the one called “Grassmann” in this submission. The manuscript does not clearly establish its relationship to these studies, and experimental comparisons may also be necessary.

References

  • Vishwanathan et al., Binet-Cauchy kernels on dynamical systems and its application to the analysis of dynamic scenes, IJCV, 2007.
  • Fujii et al., Koopman spectral kernels for comparing complex dynamics: Application to multiagent sport plays, ECML-PKDD, 2017.
  • Ishikawa et al., Metric on non-linear dynamical systems with Perron-Frobenius operators, NeurIPS, 2018.
  • Shiraishi et al., Neural decoding of electrocorticographic signals using dynamic mode decomposition, Journal of Neural Engineering, 2020.
Questions for authors
  • The method assumes that the spectral radius of the matrix A is less than 1. Could the matrix A itself contain non-decaying modes with |\lambda|=1 in practice? The external forcing term By_t can allow trajectories to continue fluctuating, but even when forcing is explicitly modeled, how is excluding such modes justified in applications involving climate data?
  • Section 3.2 uses PCA for dimensionality reduction. Is this a common procedure for climate data? To my knowledge, state-space models are often used to distinguish between states and observations; for example, Vishwanathan et al. (2007) and Fujii et al. (2017) use state-space formulations.
  • How are the choices of 20 and 200 components justified for comparing dynamics in this study?
  • Why is ARI not used for evaluation in Section 3.2?
  • Assuming comparisons such as those in Section 3.2 are commonly performed in climate science, what existing methods are available? Are they based on modal decomposition and subspace methods? Could time-series statistics, such as autocorrelation, be used for comparison without estimating a linear operator A in practice?
5 comments

Thank you for finding the framework novel and our argument convincing that a single Grassmannian subspace loses the persistence ordering a flag preserves.

W1. In climate evaluation, our contribution is the distance itself. Our distance reduces each comparison of two decompositions to one real number so that whole ensembles can be compared at once. The grouping of members by source model was not guaranteed, because the ESMs target the same system (i.e., are intended to be indistinguishable), and the resulting estimates come from finite-record statistics of each member with its own internal variability. Grouping by source models thus shows that model differences are also represented in their dynamic mode decompositions. We rewrote Sec. 3.2 around two questions. (Q1) Are the persistent modes of variability a structural property of each model? For the temperature variables, the models are distinguishable (Fig. 4b), so the subspaces spanned by their persistent modes, and the order of these modes, are a structural property of each model. For SLP, internal variability blurs the structural differences without erasing them. The same analysis shows what one member cannot: on SAT, the fast modes of CESM2 decay within about one fit step, so their order is poorly determined and changes across members (Fig. 11). (Q2) How do the persistent modes of each model differ from those of the observed climate? Each member gets one distance to 20CRv3, so every model gets a median and a spread, and no model is closest on every variable (Fig. 7). Members of one model also differ in their distance to 20CRv3, for OSAT, by about as much as the models' medians differ (Fig. 7), so no model is systematically closer to the reanalysis. These flags carry physical content (Fig. 10): the 20CRv3 OSAT flag places two slow real modes, combining Atlantic and Pacific multidecadal variability, before the ENSO-like pair, whereas most ESM members begin with that pair, consistent with the underestimated multidecadal variability in models (Kravtsov et al., 2024).

W2. We added SGOT as a baseline. The flag has a higher silhouette and ARI than SGOT on every climate variable (Fig. 4b), and SGOT does not react when decay times cross (Fig. 3). The introduction now positions our work against the Binet-Cauchy, Koopman spectral, and Perron-Frobenius methods. These do not take a fitted operator as input: Binet-Cauchy kernels compare systems through trajectories and initial conditions, and the other two require every realization to be fitted in one shared RKHS feature space. Thus, they cannot be directly applied to our setting. Like SGOT, our distance needs only each realization's own fitted AA, from any estimator; unlike SGOT, it compares invariant subspaces in persistence order (Sec. 3). We cite Shiraishi et al. (2020) as the precedent for our Grassmannian baseline.

Q1. We removed the assumption: the flag uses only the ordering of ∣λ∣|\lambda|, so it admits marginal and growing modes, and Prop. 2 holds without a bound on the spectral radius. In our climate DMDc fits, forcing enters through BytB y_t, so AA models approximately stationary internal variability. We have verified that fitted eigenvalue magnitudes are all less than 1 (App. H).

Q2. Yes, EOF truncation is the standard first step for gridded climate fields (Lorenz, 1956; Hannachi et al., 2007). A state-space estimator could replace DMDc, since the flag needs only each fitted AA (App. H).

Q3. Both come from the DMDc fits of Mankovich et al. (2026). The value 20 for OSAT and SAT follows ForceSMIP (Wills et al., 2026), and SLP needs more components (200) because it is noisier and contains a weaker forced response (App. H).

Q4. We now report ARI (Fig. 4b). The flag is comparable to the Grassmannian on OSAT and SLP. On SAT, it scores lower because it splits CESM2, whose levels at small decay times are closely spaced and change order across members (Fig. 11), the regime excluded by Prop. 2. We report this as a finding about CESM2 and as a limitation (Sec. 4).

Q5. Climate models are commonly compared with reanalyses through summary statistics: lagged autocorrelations, power spectra, and pattern correlations of fields or leading EOFs, many of which are automated in ESMValTool (Eyring et al., 2016). These statistics need no fitted AA, but most treat one location at a time. Specifically, a grid-point autocorrelation is a single diagonal entry of the lagged covariance matrix, so the off-diagonal entries, which carry the coupling across locations, are left out. Retaining them amounts to fitting AA by least squares, whose eigenvectors are the dynamic modes (Penland and Magorian, 1993). Mode-by-mode comparisons keep this coupling, but they handle one realization at a time (Mankovich et al., 2026). The flag distance reduces each comparison to a single number so that we can compare whole ensembles. We added this discussion to Sec. 1.

Thank you for your responses and for revising the manuscript.

W1-(Q1). The authors mention that ``The grouping of members by source model was not guaranteed'' and introduce (Q1) to ask whether persistent modes are structural properties of each model. This raises a concern about the evaluation in Fig. 4(b). If even an appropriate method need not group members by model, stronger agreement with model labels does not necessarily indicate a better method for comparing dynamics. Indeed, the revised manuscript appears to describe the low ARI for SAT as both a limitation and a finding about CESM2. Such framing calls into question the grounds for the comparisons in Fig. 4(b) and leaves room for alternative interpretations of the low scores of other methods. If the purpose is still to recover the model labels, the original concern remains: it is unclear how recovering these groupings contributes to climate research beyond validating the proposed method.

W1-(Q2). The additional analysis selects representative members using the proposed method and compares their modes with those of the reanalysis. It seems that Fig. 10 and its connections to the previous study (Kravtsov et al., 2024) offer an interesting interpretation. However, there is no comparison with selecting a representative member using other distances, such as the Grassmann distance, and then performing the same mode analysis. The statistics mentioned in the response to Q5 could also be used to define distances and select a representative member. That response explains the limitations of grid-point autocorrelation, but I believe that other statistical comparisons can capture relationships between locations and can also summarize a comparison as a single number. Such existing methods or their possible combinations might yield similar findings. Overall, it is still unclear what distinctive findings or interpretations the proposed method provides in this climate analysis.

W2. I thank the authors for adding SGOT as a baseline, which alleviates my original concern. However, SGOT often performs worse than Raw A and scores similarly to Euclidean under observation or process noise in Fig. 2. This may reflect limitations of SGOT, but it also raises concerns about how the synthetic data and settings of SGOT affect the results. Moreover, the original manuscript stated that SGOT shares the properties, such as gauge invariance, used to motivate the proposed method and distinguish it from other methods. The revision excluded hierarchy from this claim, but SGOT still performs poorly despite sharing the remaining properties. This leaves it unclear how these properties contribute to practical advantages in comparing dynamics. The additional references help clarify the positioning of the manuscript. Even if these methods cannot take a fitted A directly as input, the trajectories used to estimate A are available, so I think comparisons are possible. I do not consider these comparisons essential in this research, but it would be more appropriate to state that the experiments focus on methods using fitted A, rather than stating that ``they cannot be directly applied to our setting,'' as in the response.

Q1. While removing the restriction on the spectral radius seems preferable, Prop. 2 has also been modified so that the final bound additionally depends on r_{\max}. I feel the response is somewhat imprecise on this point.

Q4. I appreciate the results of ARI added in response to my question. However, if recovering the model labels is taken as the objective, Fig. 4(b) shows no clear advantage over Grassmann in terms of ARI, although the proposed method shows higher silhouette scores.

We thank xTow for their reply and thoughtful comments. We have updated the manuscript (new changes in blue).

W1-(Q1). We agree that recovering model labels does not make one distance better than another, and we have reduced the manuscript’s emphasis on it. Each distance compares a different property of the fitted operators: the Grassmannian compares the span of the leading modes, the flag compares these subspaces nested in persistence order, and the Euclidean baseline compares sign-sensitive eigenvectors. The contributions (Sec. 1) now present the flag as a complement to these baselines for questions of persistence order.

That being said, recovering the model labels still serves two purposes. First, it is a sanity check on the distance: where model structure exists, a distance should find it. Second, for climate evaluation, it answers (Q1): do the persistent modes and their order differ more between ESMs than between members of one ESM? If they do, a single member can stand for its model, which is the assumption behind mode-by-mode studies of one example member per model (Mankovich et al., 2026). They do for both temperature variables; for SLP, internal variability blurs this without erasing it (Sec. 3.2).

The ARI also depends on the clustering algorithm, so we recomputed it with average and Ward linkage on the same distance matrices (new table in App. I). Under Ward linkage, the flag and Grassmannian distances both recover all four ESMs exactly on every variable, so the lower complete-linkage scores come from the linkage rather than from the distances. Two measures in the manuscript do not depend on a clustering algorithm at all: the silhouette and the share of variance that the model labels explain in the classical MDS embedding (App. J).

The CESM2 case shows both sides of clustering. The flag detects that the order of CESM2’s fast SAT modes changes across members (Fig. 12), which the Grassmannian, by pooling levels 2 and 3, cannot see. Complete linkage exposes this as a split and Ward linkage does not, so the clustering depends on the algorithm, but where the split appears it shows which members to inspect. The manuscript now describes the split as linkage-dependent and no longer lists it as a limitation (Sec. 4).

W1-(Q2). We made the suggested mode maps for medoids selected with the Grassmannian distance (new figure in App. I). The Grassmannian picks different members, although the flag medoid ranks second under it in three of four ESMs. Every medoid keeps the signature C R R and nearly the same leading ENSO-like pair. Only the faster real modes differ, and in MIROC6 modes 3 and 4 swap order. The finding of Fig. 10 therefore does not depend on how the medoid is chosen. This is expected: our Grassmannian baseline is the last level of the flag, so the squared flag distance is the squared Grassmannian distance plus the terms of the earlier levels (Sec. 3). Both distances thus compare the span of the leading modes, and the flag also compares their persistence order. In Fig. 10, the ESMs and 20CRv3 differ in this order (C R R against R R C), which the Grassmannian cannot register because it sees only the final span. We also agree that statistics retaining spatial coupling can be summarized as one number and may support related conclusions; Sec. 4 lists this comparison as future work.

W2. We ran SGOT with the POT defaults (except r = 1) without tuning, and it did not perform well in our setting. A likely reason is that our synthetic clusters never differ in eigenvalues and eigenvectors at the same time, which may not suit SGOT’s joint matching of eigenpairs. Sec. 3 now states both points. We agree that gauge invariance alone gives no practical advantage, since SGOT shares it and still scores lower. What separates the flag is that it matches subspaces level by level in persistence order. Fig. 3 isolates this property: only the flag reacts when two decay times cross. We also revised “cannot be directly applied” to say that our experiments focus on methods that take a fitted A as input, and trajectory-based methods are left for future work.

Q1. The reviewer is right that our response was imprecise. Removing the spectral-radius assumption puts rmax⁡r_{\max} into the second inequality of Prop. 2, a factor the earlier version dropped by using rℓ<1r_{\ell} < 1. For stable operators, rmax⁡<1r_{\max} < 1 and the earlier bound is recovered. This covers our climate fits, whose spectral radii are less than 11 (App. H). The sentence after Prop. 2 now says so.

Q4. We agree that Fig. 4(b) shows no ARI advantage over the Grassmannian, and the contributions no longer suggest one. The new linkage table confirms that the two subspace distances are comparable in ARI. The flag differs in what it measures: the crossing experiment (Fig. 3), the CESM2 ordering (Fig. 12), and the reanalysis comparison (Fig. 10) rest on the persistence order that the flag records.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for AC & All Reviwers

We thank the reviewers for their feedback. We hope to have addressed all their concerns. All changes in the manuscript are marked by a color. The first round are in red, the second round are in blue. We directly address each weaknesses as W# and each questions as Q# numbered in the order they appear in each reviewer's initial response.

In response to their main concerns, we added experiments and clarified where our work sits in physical sciences in general and in climate science in particular.

First, our added/ changed figures and tables are:

  • SGOT as a baseline (Figs. 2–4; XTow, R6ic).
  • The observation-noise ablation at T=50T=50, and the ARI gap to Raw A for longer records (Figs. 2 and 5; gBze).
  • A plot of the spatial patterns (a.k.a. modes) corresponding to the ensemble member that is the flag medoid of each ESM (Fig. 10; XTow, gBze) and Grassmannian medoid (Fig. 11; XTow).
  • ARI for the climate experiment (Fig. 4b, bottom row; XTow, gBze). This revealed that the fast SAT modes of CESM2 change order across ensemble members (Fig. 12).
  • The explained variance from classical MDS, as a direct evaluation of the embedding (Fig. 13; gBze).
  • ARI under average and Ward linkage (Tab. 3; xTow).

Next, we reframed our contribution to physical and climate science (Contribution (3) and Sec. 3.2; XTow, gBze). Our distance evaluates the modes of variability of an ESM in a way that scales across ensemble members. We use it to answer two questions: (1) are the persistent modes of variability a structural property of each model, shared by its members, and (2) how do the persistent modes of each model differ from those of the observed climate, as represented by the 20CRv3 reanalysis? For (1), our evidence points to yes for temperature variables. For (2), the flag distance measures each member’s distance to the reanalysis, and the mode maps of representative members show where the models differ: they place an ENSO-like pair before the slow multidecadal modes that lead in 20CRv3 OSAT.

We also position our work against the related methods suggested by XTow (Sec. 1 and Baselines in Sec. 3).

Finally, xTow’s question about the spectral radius showed us that our method is more general than we had claimed: the theory and the method apply to fitted operators with eigenvalues of any modulus (Prop. 2 and the second paragraph of Sec. 2.2).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.