Pith. sign in

REVIEW 5 major objections 6 minor 61 references

On the Optimization of Singular Spectrum Analyses: A Pragmatic Approach

T0 review · 5 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read 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…

desk verdict Useful assembly of known SSA tricks, but the phase-preservation claim rests on a single comparison against a downsampled baseline and needs stronger validation. read the letter →

arxiv 2412.17793 v1 pith:BQSO4FGZ submitted 2024-12-23 physics.geo-ph cs.PFphysics.bio-ph

classification physics.geo-phcs.PFphysics.bio-ph
keywords SingularSpectrumAnalysisrandomizedSVDHankelmatrixhierarchicalclusteringboundaryeffectspolarmotiontreetemperaturetimeseriesdecomposition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

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.

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 (5)
  1. [§3.2 and §4] 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.
  2. [§3.1] 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.
  3. [§2.3.2 and Figure 2] 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.
  4. [§2.3.3] 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.
  5. [§3.2] 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.
minor comments (6)
  1. [§4 and Appendix A] 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.
  2. [Figure 2 caption and §2.3.2] 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.
  3. [§3.2] 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.
  4. [§2.3.5] 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.
  5. [Appendix A] 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.
  6. [Figures 4–7] 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.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: pragmatic SSA is validated against independent canonical SSA; self-citations are contextual.

full rationale

The paper's central claim is that the pragmatic-SSA pipeline (mirrored edge extension, randomized SVD, 90% energy threshold, hierarchical clustering) reproduces the phase and amplitude of components obtained by canonical SSA. This is validated by direct comparison against canonical SSA, a method developed independently by Vautard and Ghil (1989) and implemented here on the same real datasets. The pragmatic pipeline's parameters (L=1100, q=1000, 90% threshold, average-linkage clustering) are fixed algorithmic choices, not fitted to minimize discrepancy with the canonical result; no parameter is tuned to the target components. The mirrored-extension step is tested in Section 3.2 and Figure 11 rather than assumed in the validation, and the phase-shift plot compares two independently computed reconstructions. The paper's self-citations (Boulé et al. 2024, Le Mouël et al. 2024, etc.) supply datasets, prior applications, and an earlier canonical-SSA figure, but these are contextual; the load-bearing benchmark, canonical SSA, is not a self-citation. The acknowledged weakness that the canonical reference was computed on data downsampled by a factor of 1000 affects the strength of the validation, not its circularity, because the two pipelines operate on different effective inputs and their agreement is not enforced by construction. No equation in the paper defines a pragmatic output in terms of the canonical output, and no fitted quantity is renamed as a prediction. Hence no circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a chain of heuristic choices (window size, rank, power iterations, energy cutoff, linkage) and on two unproven premises: that mirrored extension preserves phase, and that the 90% energy cutoff keeps the physically relevant modes. None of these are derived or given error bounds; all are validated only through two illustrative case studies.

free parameters (5)
  • Window size L = L=1100 (polar motion), L=10 hours (tree data)
    Chosen per application without a stated selection rule; affects which components are separable (Section 2.3.1).
  • Randomized SVD rank q = q=1000 (polar motion), q=300 (Figure 2)
    The rank is set ad hoc; the paper says 'we arbitrarily reduced the rank' in Section 2.3.2, and no convergence or error check is given.
  • Power iterations = 2 iterations (Figure 2), 4 iterations (Appendix code)
    Number of power iterations chosen by hand with no criterion; larger values improve accuracy but cost more compute (Section 2.3.2).
  • Energy threshold = 90% cumulative energy
    A fixed 90% cutoff is used to retain eigentriplets; no analysis shows that all physically relevant components pass this threshold (Section 2.3.3).
  • Clustering linkage and distance = Average linkage on 1 - abs(correlation)
    The distance metric is chosen once; the authors admit it 'may not be the most appropriate for our biophysical data and is unlikely to be generalisable' (Section 4).
assumptions (4)
  • domain assumption Signal components are weakly separable by SSA with the chosen window L.
    The decomposition into Chandler, annual, trend, and tidal components assumes separability, which is known to be problematic for closely spaced frequencies like the K1/S1/O1 tides (Section 2.2).
  • standard math Randomized SVD with power iterations provides accurate dominant singular subspaces for the Hankel matrices used.
    Convergence and error bounds are assumed from the randomized linear algebra literature (Nakatsukasa and Tropp 2024); the paper does not verify approximation error for its specific matrices.
  • ad hoc to paper Retaining 90% of cumulative squared singular values keeps all signal components of interest and discards only noise.
    The threshold is presented as a simple rule (Section 2.3.3) without a derivation linking energy retention to component recovery.
  • domain assumption Mirror-extension of boundary segments does not introduce phase shifts or spurious components.
    Explicitly flagged in Section 3.2 as an assumption to verify; tested on one diurnal signal only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Optimization of Singular Spectrum Analyses: A Pragmatic Approach." pith.science (2026). https://pith.science/paper/BQSO4FGZ

@misc{pith2026241217793,
  author       = {Pith},
  title        = {Pith review of: On the Optimization of Singular Spectrum Analyses: A Pragmatic Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BQSO4FGZ}},
  note         = {Machine review of arXiv:2412.17793}
}
read the original abstract

Singular Spectrum Analysis (SSA) occupies a prominent place in the real signal analysis toolkit alongside Fourier and Wavelet analysis. In addition to the two aforementioned analyses, SSA allows the separation of patterns directly from the data space into the data space, with data that need not be strictly stationary, continuous, or even normally sampled. In most cases, SSA relies on a combination of Hankel or Toeplitz matrices and Singular Value Decomposition (SVD). Like Fourier and Wavelet analysis, SSA has its limitations. The main bottleneck of the method can be summarized in three points. The first is the diagonalization of the Hankel/Toeplitz matrix, which can become a major problem from a memory and/or computational point of view if the time series to be analyzed is very long or heavily sampled. The second point concerns the size of the analysis window, typically denoted as 'L', which will affect the detection of patterns in the time series as well as the dimensions of the Hankel/Toeplitz matrix. Finally, the third point concerns pattern reconstruction: how to easily identify in the eigenvector/eigenvalue space which patterns should be grouped. We propose to address each of these issues by describing a hopefully effective approach that we have been developing for over 10 years and that has yielded good results in our research work.

Figures

Figures reproduced from arXiv: 2412.17793 by the authors.

Figure 1
Figure 1. On the right is the frequency at which a colored segment of a time signal (right column) appears in the Hankel matrix. This is shown at the top for the original signal and at the bottom after the signal has been extended by duplicating its boundary segments. 2.3.2 Randomized-SVD This first step, which increases the presence of the end segments, also increases the size of the Hankel matrix, which can be significant i… view at source ↗
Figure 2
Figure 2. illustrates these observations. Using a synthetic signal (of size N = 15, 000 points), we constructed and diagonalized the corresponding Hankel matrix for window sizes ranging from 500 to 2500. We deliberately chose small sizes for pedagogical purposes to keep the computation time below ten seconds. For the Randomized-SVD, we arbitrarily reduced the rank to and chose two iterations for the power loop. The computatio… view at source ↗
Figure 3
Figure 3. Singular values obtained after the SVD of the Hankel matrix of the polar motion time series (in black). In red, the singular values whose the cumulative sum exceeds the threshold, set to 90%. 2.3.4 Reconstruct individual eigenvectors associated with thresholded eigenvalues At this stage, the signals components corresponding to a single, and only a single eigentriplet (Ui , Σi , Vi) are recon￾structed for all eigentr… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: The m1 component of the polar motion (gray curve). At the top, overlaid in blue, is the component reconstructed using all the eigentriplets obtained after a canonical SSA. At the bottom, overlaid in red, is the component reconstructed using all the eigentriplets obtain…
Figure 5
Figure 5. Figure 5: The Chandler wobble extracted from the data shown in [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: The seasonal oscillation extracted from the data shown in [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The polar drift extracted from the data shown in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: At the top, the dendrogram of the pragmatic-SSA, which led to the results shown in Figures 4, 5, 6, and 7. The eigentriplets paired during the first step, as well as their respective components in the data space, have been colorized. At the bottom, in black, the sum of…
Figure 9
Figure 9. Figure 9: From Boule´ et al. (2024). Temporal evolution since 2023: at the top, temperatures recorded within oak 01 at 50 cm and 100 cm (red curve), as well as in the soil at 90 cm depth (black curve), and outside the tree (attached to the tree) at 75 cm (blue curve); at the bot…
Figure 10
Figure 10. Figure 10: Different diurnal components (S1, O1, and K1) extracted during the year 2023 using canonical-SSA (blue curve) and pragmatic-SSA (red curve). instantaneous phase shift (in radians) between the two pseudo-cycles obtained by the two approaches. As can be seen, this phase…
Figure 11
Figure 11. Figure 11: Zoom on the results shown in [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 58 canonical work pages

  1. [1]

    Alla, A. et J. N. Kutz (2019). Randomized model order reduction. Advances in Computational Mathematics\/ 45 , 1251--1271

  2. [2]

    Benjamin Erichson, N., S. L. Brunton, et J. Nathan Kutz (2017). Compressed singular value decomposition for image and video processing. Dans Proceedings of the IEEE International Conference on Computer Vision Workshops , pp.\ 1880--1888

  3. [3]

    Bertero, M. et E. Pike (1982). Resolution in diffraction-limited imaging, a singular value analysis. Optica Acta: International Journal of Optics\/ 29\/ (6), 727--746

  4. [4]

    Poncela, et E

    B \'o galo, J., P. Poncela, et E. Senra (2021). Circulant singular spectrum analysis: A new automated procedure for signal extraction. Signal Processing\/ 179 , 107824

  5. [5]

    Bonizzi, P., J. M. Karel, O. Meste, et R. L. Peeters (2014). Singular spectrum decomposition: A new method for time series decomposition. Advances in Adaptive Data Analysis\/ 6\/ (04), 1450011

  6. [6]

    Boul \'e , J.-B., J. d. B. d'Ars, V. Courtillot, M. G \`e ze, D. Gibert, J.-L. L. Mou \"e l, F. Lopes, A. Maineult, et P. Zuddas (2024). On the mechanism of thermal self-regulation of trees: a kind of homeothermic observation. arXiv preprint arXiv:2409.09765\/

  7. [7]

    Broomhead, D. S. et G. P. King (1986). Extracting qualitative dynamics from experimental data. Physica D: Nonlinear Phenomena\/ 20\/ (2-3), 217--236

  8. [8]

    Chaitanya, M. K. et L. D. Sharma (2024). Cross subject myocardial infarction detection from vectorcardiogram signals using binary harry hawks feature selection and ensemble classifiers. IEEE Access\/ 12 , 28247--28259

Show all 61 references
  1. [9]

    Chandler, S. C. (1891a). On the variation of latitude, i. Astronomical Journal, vol. 11, iss. 248, p. 59-61 (1891).\/ 11 , 59--61

  2. [10]

    Chandler, S. C. (1891b). On the variation of latitude, ii. Astronomical Journal, vol. 11, iss. 249, p. 65-70 (1891).\/ 11 , 65--70

  3. [11]

    Cipra, B. A. (2000). The best of the 20th century: Editors name top 10 algorithms. SIAM news\/ 33\/ (4), 1--2

  4. [12]

    Le Mou \"e l, F

    Courtillot, V., J.-L. Le Mou \"e l, F. Lopes, et D. Gibert (2022). On sea-level change in coastal areas. Journal of Marine Science and Engineering\/ 10\/ (12), 1871

  5. [13]

    Georgiev, S

    Darnell, G., S. Georgiev, S. Mukherjee, et B. E. Engelhardt (2017). Adaptive randomized dimension reduction on massive data. Journal of Machine Learning Research\/ 18\/ (140), 1--30

  6. [14]

    Le Mou \"e l, V

    Dumont, S., J.-L. Le Mou \"e l, V. Courtillot, F. Lopes, F. Sigmundsson, D. Coppola, E. P. Eibl, et C. J. Bean (2020). The dynamics of a long-lasting effusive eruption modulated by earth tides. Earth and Planetary Science Letters\/ 536 , 116145

  7. [15]

    Erichson, N. B., S. Voronin, S. L. Brunton, et J. N. Kutz (2016). Randomized matrix decompositions using r. arXiv preprint arXiv:1608.02148\/

  8. [16]

    Francis, J. G. (1961). The qr transformation a unitary analogue to the lr transformation—part 1. The Computer Journal\/ 4\/ (3), 265--271

  9. [17]

    Francis, J. G. (1962). The qr transformation—part 2. The Computer Journal\/ 4\/ (4), 332--345

  10. [18]

    Kannan, et S

    Frieze, A., R. Kannan, et S. Vempala (2004). Fast monte-carlo algorithms for finding low-rank approximations. Journal of the ACM (JACM)\/ 51\/ (6), 1025--1041

  11. [19]

    Le Mou \"e l, L

    Gibert, D., J.-L. Le Mou \"e l, L. Lambs, F. Nicollin, et F. Perrier (2006). Sap flow and daily electric potential variations in a tree trunk. Plant Science\/ 171\/ (5), 572--584

  12. [20]

    Lopes, V

    Gibert, D., F. Lopes, V. Courtillot, J.-L. Le Mou \"e l, et J.-B. Boul \'e (2024). Information theory, signal analysis and inverse problem. arXiv preprint arXiv:2408.16361\/

  13. [21]

    Korobeynikov, et A

    Golyandina, N., A. Korobeynikov, et A. Zhigljavsky (2018). Singular spectrum analysis with R . Springer

  14. [22]

    Nekrutkin, et A

    Golyandina, N., V. Nekrutkin, et A. A. Zhigljavsky (2001). Analysis of time series structure: SSA and related techniques . CRC press

  15. [23]

    Golyandina, N. et A. Shlemov (2013). Variations of singular spectrum analysis for separability improvement: non-orthogonal decompositions of time series. arXiv preprint arXiv:1308.4022\/

  16. [24]

    Cayrols, et J

    Grigori, L., S. Cayrols, et J. W. Demmel (2018). Low rank approximation of a sparse matrix based on lu factorization with column and row tournament pivoting. SIAM Journal on Scientific Computing\/ 40\/ (2), C181--C209

  17. [25]

    Hansen, P. C. (1998). Rank-deficient and discrete ill-posed problems: numerical aspects of linear inversion . SIAM

  18. [26]

    Hariprasad, M. et M. Venkatapathi (2024). Circulant decomposition of a matrix and the eigenvalues of toeplitz type matrices. Applied Mathematics and Computation\/ 468 , 128473

  19. [27]

    Fourer, F

    Harmouche, J., D. Fourer, F. Auger, P. Borgnat, et P. Flandrin (2017). The sliding singular spectrum analysis: A data-driven nonstationary signal decomposition tool. IEEE Transactions on Signal Processing\/ 66\/ (1), 251--263

  20. [28]

    Hassani, H. (2007). Singular spectrum analysis: methodology and comparison

  21. [29]

    Hassani, H., E. S. Silva, et Z. Ghodsi (2017). Optimizing bicoid signal extraction. Mathematical biosciences\/ 294 , 46--56

  22. [30]

    Jarman, A. M. (2020). Hierarchical cluster analysis: Comparison of single linkage, complete linkage, average linkage and centroid linkage method. Georgia Southern University\/ 29

  23. [31]

    Gibert, et J

    Jourde, K., D. Gibert, et J. Marteau (2015). Improvement of density models of geological structures by fusion of gravity data and cosmic muon radiographies. Geoscientific Instrumentation, Methods and Data Systems\/ 4\/ (2), 177--188

  24. [32]

    Lambeck, K. (2005). The Earth's variable rotation: geophysical causes and consequences . Cambridge University Press

  25. [33]

    Gibert, J.-B

    Le Mou \"e l, J.-L., D. Gibert, J.-B. Boul \'e , P. Zuddas, V. Courtillot, F. Lopes, M. G \`e ze, et A. Maineult (2024). On the effect of the luni-solar gravitational attraction on trees. arXiv preprint arXiv:2402.07766\/

  26. [34]

    Lopes, et V

    Le Mou \"e l, J.-L., F. Lopes, et V. Courtillot (2017). Identification of gleissberg cycles and a rising trend in a 315-year-long series of sunspot numbers. Solar Physics\/ 292\/ (3), 43

  27. [35]

    Lemmerling, P. et S. Van Huffel (2001). Analysis of the structured total least squares problem for hankel/toeplitz matrices. Numerical algorithms\/ 27 , 89--114

  28. [36]

    Wang, et Y

    Li, X., S. Wang, et Y. Cai (2019). Tutorial: Complexity analysis of singular value decomposition and its variants. arXiv preprint arXiv:1906.12085\/

  29. [37]

    Liu, Y., C. Xu, L. Qi, et Y. Li (2024). A robust non-contact heart rate estimation from facial video based on a non-parametric signal extraction model. Biomedical Signal Processing and Control\/ 93 , 106186

  30. [38]

    Courtillot, et J.-L

    Lopes, F., V. Courtillot, et J.-L. Le Mou \"e l (2022). Triskeles and symmetries of mean global sea-level pressure. Atmosphere\/ 13\/ (9), 1354

  31. [39]

    Le Mou \"e l, V

    Lopes, F., J. Le Mou \"e l, V. Courtillot, et D. Gibert (2021). On the shoulders of laplace. Physics of the Earth and Planetary Interiors\/ 316 , 106693

  32. [40]

    Lorenz, E. N. (1963). Deterministic nonperiodic flow. Journal of atmospheric sciences\/ 20\/ (2), 130--141

  33. [41]

    Markowitz, W. (1968). Concurrent astronomical observations for studying continental drift, polar motion, and the rotation of the earth. Dans Symposium-International Astronomical Union , Volume 32, pp.\ 25--32. Cambridge University Press

  34. [42]

    Martinsson, P.-G. (2019). Randomized methods for matrix computations. The Mathematics of Data\/ 25\/ (4), 187--231

  35. [43]

    Menke, W. (1984). The resolving power of cross-borehole tomography. Geophysical Research Letters\/ 11\/ (2), 105--108

  36. [44]

    Miller, G. (1974). Fredholm equations of the first kind. Numerical solution of integral equations\/ , 175--188

  37. [45]

    Murtagh, F. et P. Contreras (2012). Algorithms for hierarchical clustering: an overview. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery\/ 2\/ (1), 86--97

  38. [46]

    Nakatsukasa, Y. (2020). Fast and stable randomized low-rank matrix approximation. arXiv preprint arXiv:2009.11392\/

  39. [47]

    Nakatsukasa, Y. et J. A. Tropp (2024). Fast and accurate randomized algorithms for linear systems and eigenvalue problems. SIAM Journal on Matrix Analysis and Applications\/ 45\/ (2), 1183--1214

  40. [48]

    Noorizadegan, A., C.-S. Chen, R. Cavoretto, et A. De Rossi (2024). Efficient truncated randomized svd for mesh-free kernel methods. Computers & Mathematics with Applications\/ 164 , 12--20

  41. [49]

    Oropeza, V. et M. Sacchi (2011). Simultaneous seismic data denoising and reconstruction via multichannel singular spectrum analysis. Geophysics\/ 76\/ (3), V25--V32

  42. [50]

    Park, J., C. R. Lindberg, et F. L. Vernon III (1987). Multitaper spectral analysis of high-frequency seismograms. Journal of Geophysical Research: Solid Earth\/ 92\/ (B12), 12675--12684

  43. [51]

    Penrose, R. (1955). A generalized inverse for matrices. Dans Mathematical proceedings of the Cambridge philosophical society , Volume 51, pp.\ 406--413. Cambridge University Press

  44. [52]

    Ros, F. et S. Guillaume (2019). A hierarchical clustering algorithm and an improvement of the single linkage criterion to deal with noise. Expert Systems with Applications\/ 128 , 96--108

  45. [53]

    Santhanam, B. et P. Maragos (2000). Multicomponent am-fm demodulation via periodicity-based algebraic separation and energy-based demodulation. IEEE Transactions on Communications\/ 48\/ (3), 473--490

  46. [54]

    Do g an, et \.I

    Sara c li, S., N. Do g an, et \.I . Do g an (2013). Comparison of hierarchical cluster analysis methods by cophenetic correlation. Journal of inequalities and Applications\/ 2013 , 1--8

  47. [55]

    Stoyko, A. (1968). Mouvement seculaire du pole et la variation des latitudes des stations du sil. Dans Symposium-International Astronomical Union , Volume 32, pp.\ 52--56. Cambridge University Press

  48. [56]

    Rosenbach, et M

    Tamascelli, D., R. Rosenbach, et M. B. Plenio (2015). Improved scaling of time-evolving block-decimation algorithm through reduced-rank randomized singular value decomposition. Physical Review E\/ 91\/ (6), 063306

  49. [57]

    Gallet, F

    Usoskin, I., Y. Gallet, F. Lopes, G. Kovaltsov, et G. Hulot (2016). Solar activity during the holocene: the hallstatt cycle and its consequence for grand minima and maxima. Astronomy & Astrophysics\/ 587 , A150

  50. [58]

    Vandebril, et K

    Vannieuwenhoven, N., R. Vandebril, et K. Meerbergen (2012). A new truncation strategy for the higher-order singular value decomposition. SIAM Journal on Scientific Computing\/ 34\/ (2), A1027--A1052

  51. [59]

    Vautard, R. et M. Ghil (1989). Singular spectrum analysis in nonlinear dynamics, with applications to paleoclimatic time series. Physica D: Nonlinear Phenomena\/ 35\/ (3), 395--424

  52. [60]

    Rogosin, et G

    Vitali, G., S. Rogosin, et G. Baldoni (2019). Climate change and grain production fluctuations. arXiv preprint arXiv:2002.07039\/

  53. [61]

    write newline

    " write newline "" before.all 'output.state := FUNCTION article output.bibitem format.authors "author" output.check author format.key output output.year.check new.block format.title "title" output.check new.block crossref missing format.jour.vol output format.article.crossref ...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.