{"id":"521117d9-681c-43fd-ad0a-b9a51250c318","arxiv_id":"2412.17793","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A pipeline combining mirrored signal extension, randomized SVD, energy thresholding, and hierarchical clustering reproduces classical SSA components on geophysical and biophysical time series without downsampling.","lead":"This paper proposes a practical recipe to make Singular Spectrum Analysis fast on very long time series: mirror-extend the signal, use a randomized SVD, keep the top-energy components, and cluster them automatically. The authors show the recipe matches classical SSA on polar-motion and tree-temperature records, in phase and amplitude, without downsampling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phase-preservation claim rests on a single comparison against a downsampled canonical SSA, which cannot serve as ground truth; the mirrored-extension assumption is therefore untested.","rationale":"The reader's weakest assumption identified the mirrored-extension phase-shift concern, and I agree that this is load-bearing. However, my analysis sharpens the issue: the single verification in Figure 11 is not a valid falsification because it compares pragmatic-SSA against a canonical-SSA baseline that is itself run on 1000x-downsampled data, so agreement with that baseline cannot establish absence of phase bias relative to the true signal. In addition, the central equivalence claim is supported only by visual overlay and no quantitative agreement metric, and the Chandler component discrepancy is acknowledged but not resolved. These issues together mean the paper's central claim is plausible but not yet demonstrated. The proposed synthetic test with known phases provides a direct, quantitative check of the mirror-extension assumption and of the full-resolution versus downsampled comparison. Since the reader already returned a CONDITIONAL verdict, my stress-test does not change that verdict; it reinforces the conditionality with a more precise falsification target.","tokens_in":18291,"tokens_out":4492,"duration_ms":49481,"concrete_test":"Generate a synthetic 1 Hz signal of length comparable to the tree segment (about 3 months) as a sum of known diurnal sinusoids (S1, K1, O1 with known phases and amplitudes) plus colored noise. Run (a) pragmatic-SSA with mirrored extension, (b) pragmatic-SSA without extension, and (c) canonical SSA on the same full-resolution series and on the 1000x-downsampled series. Estimate the instantaneous phase of the extracted diurnal component via Hilbert transform and compute mean absolute phase error against the known true phase over the central 80% of the record. Reject the phase-preservation claim if the mirrored-extension error is significantly larger than the no-extension error or than the downsampled-canonical error across more than 10 noise realizations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that pragmatic-SSA extracts signal components comparable in phase and amplitude to canonical SSA. Section 3.2 explicitly asks whether mirrored edge extension introduces phase shifts and answers with Figure 11, which overlays pragmatic-SSA against canonical SSA run on data downsampled by a factor of 1000 (Section 4). Because the reference itself is an approximation on different information, a near-zero phase difference only shows that the two approximations agree; it cannot detect a common bias relative to the true signal, and it cannot isolate the effect of mirrored extension. The polar-motion comparison (Figures 4 to 7) is also purely visual, admits an unresolved Chandler-period discrepancy after 2000, and compares methods run on different effective datasets (canonical SSA with L=2900 on 3577 points; pragmatic SSA with L=1100 on the extended series). Thus the most load-bearing condition for the central claim, phase and amplitude equivalence, is not actually established by a falsifiable test. The mirror-extension step is the least secure ingredient because the validation protocol cannot rule out a phase bias that affects both methods or only the pragmatic pipeline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a \"pragmatic SSA\" pipeline intended to make Singular Spectrum Analysis feasible on long, high-sampling-rate time series without downsampling. The pipeline combines four ingredients: (1) extending the signal by mirroring boundary segments of length L to reduce boundary effects in the Hankel matrix; (2) replacing the exact SVD with a randomized SVD of fixed rank q and a small number of power iterations; (3) retaining eigentriplets whose cumulative squared singular values reach a 90% energy threshold; and (4) reconstructing each retained eigentriplet individually, removing the extension, and grouping the reconstructed components by average-linkage hierarchical clustering of a correlation-derived distance. The method is demonstrated on two real datasets: the m1 component of polar motion (3577 points) and tree temperature recordings at 1 Hz over roughly a year, which the authors state would be on the order of π×10^7 points. The central claim is that the pragmatic pipeline recovers the same principal components—in phase and amplitude—as canonical SSA, while avoiding downsampling and completing within reasonable time on standard hardware.","tokens_in":18467,"tokens_out":4280,"duration_ms":39049,"significance":"If the central claim were convincingly validated, the paper would offer a practical and useful contribution: canonical SSA is computationally expensive for long series, and manual eigentriplet grouping is a recognized bottleneck. The use of randomized SVD and automated clustering is a sensible direction, and the MATLAB code in Appendix A supports reproducibility. The comparison against canonical SSA is an appropriate external benchmark and is not circular. However, as it stands, the validation is largely qualitative and uncontrolled: the reference canonical SSA is run on downsampled data, several key parameters are selected arbitrarily without sensitivity analysis, and the mirror-extension phase-preservation assumption is tested only once against that same downsampled reference. These weaknesses are load-bearing for the paper's central claim and need to be addressed before the contribution can be assessed as established.","major_comments":[{"comment":"The validation of phase and amplitude equivalence is carried out against canonical SSA applied to data downsampled by a factor of 1000 (one point every 0.2 hours), as the paper itself states in §4: \"Because of this downsampling, in principle, the two methods do not analyze the same information.\" Since the reference is itself an approximation on coarser information, the near-perfect overlay in Figure 11 cannot establish that pragmatic-SSA preserves the phase and amplitude of the true signal; it only establishes agreement between two approximations. A common bias, such as an undetected phase shift from the mirror extension, would be invisible in this comparison. The central claim therefore needs a validation against a known-truth synthetic signal or a higher-fidelity reference, not only against a downsampled canonical SSA.","section":"§3.2 and §4"},{"comment":"The polar-motion comparison is not a controlled comparison: canonical SSA uses L=2900 on the 3577-point series, while pragmatic SSA uses L=1100 on the mirror-extended series, and the two arms retain different numbers of components (first 10 eigentriplets versus 38 eigentriplets grouped into 23 clusters). The comparison is purely visual, with no quantitative error metric or phase statistic, and the Chandler component shows a significant post-2000 discrepancy that the paper leaves unresolved (\"Whether this is correct or not is open to interpretation\"). Given that this example is offered as evidence for \"amplitude and phase preserved,\" the lack of any quantitative measure and the unexplained discrepancy substantially weaken the claim.","section":"§3.1"},{"comment":"The randomized-SVD parameters are chosen arbitrarily and no sensitivity analysis is provided. The main text states \"For the Randomized-SVD, we arbitrarily reduced the rank to and chose two iterations for the power loop,\" while the Figure 2 caption says \"The chosen rank (q) is arbitrarily set to L = 300 points\"—the text is missing the rank value and the caption appears inconsistent with the text. Since the accuracy of the randomized SVD and the resulting component separation depend directly on q and the number of power iterations, the arbitrary selection of these parameters without any robustness check leaves the reliability of the pipeline unsupported.","section":"§2.3.2 and Figure 2"},{"comment":"The 90% cumulative-energy threshold is introduced without justification or sensitivity analysis. The paper asserts that this retains all signal components of interest and discards only noise, but no evidence is shown that the reconstruction is stable with respect to the threshold, nor how the number of retained eigentriplets or the resulting components vary as the threshold is changed. Because this threshold directly determines the set of components that are reconstructed and clustered, its arbitrary choice is load-bearing for the claimed agreement with canonical SSA.","section":"§2.3.3"},{"comment":"The mirror-extension phase-preservation assumption is explicitly acknowledged as an assumption (\"Intuitively, it seems reasonable to assume that it does not [introduce phase shifts]\") and is tested only once, on a single diurnal temperature component, and against the same downsampled reference identified above. This is a single qualitative test, not a systematic check across components, segments, or signal types. Given that mirrored edge extension is a core novelty of the proposed pipeline, a targeted test on synthetic signals with known phase is needed to rule out a phase bias, especially one that affects both methods equally and would therefore be invisible in the current comparison.","section":"§3.2"}],"minor_comments":[{"comment":"There are several typos and formatting issues: \"denationalization\" should be \"diagonalization\" (appears twice in §4 and once in Appendix A), \"Setion 2.1\" should be \"Section 2.1\", and the sentence \"each segment is could be counted up to 20 times per column\" in §2.3.1 is ungrammatical.","section":"§4 and Appendix A"},{"comment":"The Figure 2 caption states \"The chosen rank (q) is arbitrarily set to L = 300 points,\" but the main text says \"we arbitrarily reduced the rank to and chose two iterations\" with the rank value missing. Please clarify the exact rank and ensure consistency between text and caption.","section":"Figure 2 caption and §2.3.2"},{"comment":"The statement \"we did not downsample the data\" in §3.2 is followed by \"we performed SSA on segmented portions of the data due to important measurement interruptions.\" It would help to state explicitly whether each analyzed segment is continuous at 1 Hz and to give the exact length of the longest segment, which is later reported as about π/4×10^7 seconds.","section":"§3.2"},{"comment":"The description of the distance used for clustering is unclear: \"the average distance of each component to the correlation matrix of all the components\" is not a well-defined phrase; the formal definition in equation (6) is clearer, but the text introducing it should be revised.","section":"§2.3.5"},{"comment":"The MATLAB code lists calls to ssaReconstruct and ssaClustering but does not provide their implementations, and the sentence \"Obviously, for a rank equal to , the approach loses its usefulness\" has a missing value. Please either include the missing functions or state where they can be obtained.","section":"Appendix A"},{"comment":"Color coding is inconsistent across figures: in Figure 4 the blue curve is called canonical-SSA and the red curve pragmatic-SSA, but in Figures 5–7 the blue curve is canonical-SSA and the red curve is pragmatic-SSA. Please standardize the color scheme and the caption descriptions.","section":"Figures 4–7"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real computational bottleneck and the proposed pipeline is plausible, but the current validation is largely anecdotal: the reference itself is computed on downsampled data, the key parameters are chosen arbitrarily, and the mirror-extension assumption is tested only once. I would recommend requiring a synthetic-signal validation with known ground truth, a sensitivity analysis for L, q, power iterations, and the energy threshold, and quantitative phase/amplitude error metrics before publication. The MATLAB code and the use of an external benchmark are strengths that should be preserved in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nQuick take on arXiv:2412.17793. The paper assembles four known SSA ingredients — mirrored edge extension, randomized SVD, cumulative-energy thresholding, and hierarchical clustering of individually reconstructed components — into a single pipeline called pragmatic-SSA. That assembly is useful for practitioners: it lets you run SSA on long, high-rate series without downsampling, and automates the eigentriplet grouping step. The two case studies are real data, and the code is in the appendix.\n\nWhat the paper does well: it is clear about what is standard and what is new. The authors openly acknowledge the unresolved Chandler-period discrepancy and that the clustering distance may not generalize. The polar-motion decomposition matches canonical SSA reasonably well, and the computational savings are obvious.\n\nWhere it is soft: the validation is essentially qualitative. The main comparison is against canonical SSA run on downsampled data (factor 1000 for the tree example), so the two methods do not see the same information. More importantly, the phase-preservation claim — that mirrored extension does not shift phases — is load-bearing and is tested only once, on a single diurnal temperature component. That test cannot isolate the mirror-extension effect because both arms might share a common bias relative to the true signal; a zero phase difference between the two approximations only shows they agree with each other, not that either is unbiased. The parameter choices (L, rank, power iterations, energy threshold) are arbitrary and there is no sensitivity analysis. None of this is hidden; the paper admits the limitations, but the central claim is stronger than the evidence supports.\n\nThe stress-test note is on target. The assumption that mirror extension preserves phase is stated as \"intuitively reasonable\" (Section 3.2) and then verified once. That is not enough for a method whose whole point is to avoid downsampling.\n\nWho it is for: applied time-series analysts who want a practical SSA workflow and can tolerate some uncertainty about the phase-preservation guarantee. It deserves a serious referee — the assembly is useful, the code is there, and the limitations are honestly flagged — but a referee should ask for a controlled comparison on the same data, quantitative agreement metrics, and sensitivity analysis. If those are added, the paper would be solid.\n\nI would bring it to reading group as an example of a methods paper with an honest but under-powered validation. My verdict: conditional, leaning toward accept with major revision.\n\nBest,\n[You]","headline":"Useful assembly of known SSA tricks, but the phase-preservation claim rests on a single comparison against a downsampled baseline and needs stronger validation.","tokens_in":19040,"tokens_out":3323,"would_cite":true,"duration_ms":29942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a four-step \"pragmatic-SSA\" pipeline—mirror extension, randomized SVD, energy thresholding, and data-space clustering—extracts the same geophysical and biophysical modes as canonical SSA while running on…","keywords":["Singular Spectrum Analysis","randomized SVD","Hankel matrix","hierarchical clustering","boundary effects","polar motion","tree temperature","time series decomposition"],"falsifier":"Generate a synthetic series with two closely spaced oscillations of known frequencies, amplitudes, and phases, sampled at 1 second for several months. Run pragmatic-SSA (mirror extension, randomized SVD, 90% threshold, average-linkage clustering) and compare each reconstructed component's instantaneous phase and amplitude against the true generating functions. If the measured phase error is not near zero across the full record, or if it changes when the mirror-extension length $L$ is varied, the assumption that mirrored edges do not bias reconstruction phases is falsified.","tokens_in":18037,"feed_emoji":"📊","tokens_out":6139,"duration_ms":52691,"temperature":0.7,"pith_summary":"Singular Spectrum Analysis (SSA) decomposes a time series into trends and pseudo-cycles by embedding it in a Hankel matrix and applying SVD, but the matrix diagonalization becomes too memory-hungry and slow for long, densely sampled geophysical and biophysical records. The paper proposes a four-step pragmatic-SSA pipeline—mirror-extending the signal ends, replacing exact SVD with randomized SVD, retaining the singular values that cumulatively hold 90% of the energy, and grouping individually reconstructed components with average-linkage hierarchical clustering in the data space. The claim is that this pipeline extracts the same dominant components, with the same phases and amplitudes, as canonical SSA, while working on the full-resolution signal without downsampling and in feasible time on a standard computer. The paper supports the claim on two real datasets: the m1 component of polar motion, where the first clusters reproduce the Chandler wobble, the annual oscillation, and the Markowitz–Stoyko drift; and tree temperature recordings, where the diurnal tidal components match canonical SSA with an instantaneous phase shift near zero.","feed_headline":"Randomized SSA matches classic signal splits, no downsampling","feed_subtitle":"A four-step pipeline preserves tide phases in tree data and Chandler wobble in polar motion, on standard hardware.","key_machinery":"The load-bearing mechanism is the four-step pipeline, each step targeting one bottleneck. Mirrored edge extension duplicates and flips the first and last $L$ points of the signal before embedding, so the boundary segments appear as often in the Hankel matrix as interior segments, mitigating the boundary effect. Randomized SVD then projects the Hankel matrix onto a random subspace of rank $q$, applies power iterations, orthonormalizes with QR, and computes an exact SVD of the reduced $q \\times K$ matrix, lowering complexity from $O(LK^2)$ to $O(LKq)+O(q^2(L+K))$. A 90% cumulative-energy threshold on the squared singular values selects the retained eigentriplets. Each retained eigentriplet is reconstructed separately in the data space, and average-linkage hierarchical clustering on the distance matrix $1-|\\rho|$ groups components that belong to the same oscillation, so grouping happens in data space rather than in the abstract eigenvector space.","core_discovery":"The central discovery is that the computational bottlenecks of SSA can be bypassed without sacrificing the signal content that matters. On the polar-motion series, a canonical SSA with window $L = 2900$ recovers the Chandler wobble, annual oscillation, and drift in eigentriplets $\\{1,2\\}$, $\\{3,4\\}$, and $\\{5\\}$; the pragmatic pipeline with $L = 1100$, rank $q = 1000$, and a 90% cumulative-energy threshold returns 38 eigentriplets grouped into 23 clusters whose first three clusters reproduce the same three components almost identically. On the tree data, for a segment of about $\\pi/4 \\times 10^7$ points at 1-second sampling, pragmatic-SSA without downsampling extracts the S1, K1, and O1 diurnal temperature oscillations that track the canonical-SSA curves in phase and amplitude, with a measured instantaneous phase shift close to zero over the whole overlap. The paper states this as evidence that the mirrored extension does not introduce phase distortion and that the truncated randomized decomposition, energy threshold, and data-space clustering preserve the physically meaningful components.","pith_inferences":["If the phase-neutrality of mirror extension holds generally, the same pipeline could be applied to streaming records: each new block could extend the analyzed window without re-embedding from scratch, though the paper does not test this.","The 90% threshold and average-linkage distance are heuristics; for signals with closely spaced spectral peaks, a lower threshold or a different linkage might separate modes better, and the paper itself notes the distance metric may not generalize.","A natural stress test would be a synthetic benchmark with known phases and two close frequencies, comparing pragmatic-SSA against canonical SSA on the same record; the paper's tree example is the closest existing analog and shows near-zero phase shift only for one diurnal component.","Because clustering operates on reconstructed components, the method could double as a separability diagnostic: if a dendrogram merges components across known physical modes, that flags weak separability in the data."],"forward_implications":["Full-resolution records of the size of a year of 1-second tree measurements (about $\\pi \\times 10^7$ points) can be analyzed without downsampling; the paper demonstrates this on a segment of roughly $\\pi/4 \\times 10^7$ points.","The dendrogram produced by hierarchical clustering makes the grouping of eigentriplets explicit and auditable, so a user can see which components merge at each step.","For long geophysical series, the first clusters reproduce the main physical modes—Chandler wobble, annual oscillation, polar drift—in phase and amplitude comparable to canonical SSA.","The 90% energy threshold automatically discards low-energy components, replacing manual selection of how many eigentriplets to keep."],"supporting_citations":[{"why":"Reference work defining the canonical four-step SSA procedure that pragmatic-SSA modifies.","marker":"Golyandina et al. (2018)"},{"why":"Introduces the trajectory-matrix embedding that underlies the Hankel construction.","marker":"Broomhead et King (1986)"},{"why":"Established SSA for paleoclimate time-series analysis, the geophysical context the paper extends.","marker":"Vautard et Ghil (1989)"},{"why":"Provides the classical SVD algorithm that randomized SVD replaces for large Hankel matrices.","marker":"Golub and Reinsch 1971"},{"why":"Supplies the random-projection low-rank factorization foundation for the randomized-SVD step.","marker":"Frieze et al. (2004)"},{"why":"Tree thermal self-regulation dataset and baseline canonical-SSA results used for the second comparison.","marker":"Boulé et al. (2024)"}],"fun_headline_variants":["Randomized SSA matches classic splits, no downsampling","Pragmatic SSA: same signals, far less memory","Faster SSA without downsampling: a pragmatic pipeline","SSA optimization: random tricks yield classic results","Efficient SSA handles long series on standard hardware"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that flipping and duplicating the signal's end segments does not shift the phase of the extracted oscillations; the paper tests it on a single diurnal temperature component and otherwise takes it as intuitively reasonable.","fun_headline_variants_meta":{"raw":{"variants":["Randomized SSA matches classic splits, no downsampling","Pragmatic SSA: same signals, far less memory","Faster SSA without downsampling: a pragmatic pipeline","SSA optimization: random tricks yield classic results","Efficient SSA handles long series on standard hardware"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2712,"prompt_tokens":1027,"completion_tokens":1685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1604}},"tokens_in":643,"tokens_out":1685,"duration_ms":12455,"temperature":1.0,"reasoning_tokens":1604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:21.847848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a synthetic series with two closely spaced oscillations of known frequencies, amplitudes, and phases, sampled at 1 second for several months. Run pragmatic-SSA (mirror extension, randomized SVD, 90% threshold, average-linkage clustering) and compare each reconstructed component's instantaneous phase and amplitude against the true generating functions. If the measured phase error is not near zero across the full record, or if it changes when the mirror-extension length $L$ is varied, the assumption that mirrored edges do not bias reconstruction phases is falsified.","supporting_citations":[{"cited_title":"Korobeynikov, et A","cited_arxiv_id":null,"evidence_quote":"Reference work defining the canonical four-step SSA procedure that pragmatic-SSA modifies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the trajectory-matrix embedding that underlies the Hankel construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Established SSA for paleoclimate time-series analysis, the geophysical context the paper extends."},{"cited_title":"Kannan, et S","cited_arxiv_id":null,"evidence_quote":"Supplies the random-projection low-rank factorization foundation for the randomized-SVD step."}],"review_version":1}