Pith. sign in

REVIEW 4 major objections 5 minor 26 references

SVD-Based Graph Fractional Fourier Transform on Directed Graphs and Its Application

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Two new SVD-based graph fractional Fourier transforms for directed Cartesian product graphs concentrate the energy of strongly correlated spatiotemporal signals in low frequencies and outperform prior fractional Fourier and Fourier…

desk verdict The core flaw is that fractional powers of the SVD factors need not be real orthogonal, so the directed transforms and truncation theorems don't follow as stated; the undirected reduction is correct and the idea is repairable. read the letter →

arxiv 2506.03925 v2 pith:JKMYS3RI submitted 2025-06-04 eess.SP

classification eess.SP
keywords graphsignalprocessingfractionalFouriertransformdirectedsingularvaluedecompositionCartesianproductdenoisingspatiotemporalsignals
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

This paper tries to establish that fractional-order Fourier analysis can be carried out on signals that live on Cartesian products of directed graphs, a natural setting for time–vertex or multi-sensor data such as temperature records. It defines two graph fractional Fourier transforms, $F^\alpha_\boxtimes$ and $F^\alpha_\otimes$, from singular value decompositions of fractional Laplacian matrices, and proves truncation bounds saying that when the fractional Laplacian acts mildly on a signal, almost all of the signal's energy sits in the lowest frequency components. The paper reports that on hourly temperature data from 32 weather stations, $F^\alpha_\otimes$ concentrates about 99.6% of the signal energy in the first 40 of 768 frequencies and yields higher denoising signal-to-noise ratios than an existing directed-graph fractional Fourier transform and than the non-fractional SVD-based graph Fourier transforms. If correct, the construction gives a computationally cheaper, explicitly bounded way to analyze strongly correlated spatiotemporal graph signals in the fractional domain.

What carries the argument

The engine of the paper is the fractional Laplacian together with its singular vectors. Starting from $L_l = U_l\Sigma_l V_l^T$, the paper sets $P_l=U_l^\alpha$, $Q_l=V_l^\alpha$, and $R_l=\Sigma_l^\alpha$; these $P,Q$ pairs are the building blocks for both transforms. $F^\alpha_\boxtimes$ uses the SVD of the Kronecker-sum fractional Laplacian $L^\alpha_\boxtimes$, so its frequency axis is the sorted fractional singular values $r_k$ of the whole product graph, while $F^\alpha_\otimes$ uses the Kronecker products $P_1\otimes P_2$ and $Q_1\otimes Q_2$, so its two-dimensional frequencies are pairs $(r_{1,i},r_{2,j})$ of factor fractional singular values. The truncation theorems carry the argument: the energy outside the first $\Omega$ components is controlled by the norms $\|L^\alpha x\|$ and $\|(L^\alpha)^T x\|$ divided by the cutoff frequency, which is the quantitative statement that small fractional-Laplacian action means low-frequency concentration. The inverse transforms and norm preservation are what make the bases usable for reconstruction and denoising.

What would settle it

Use the same algorithm on temperature data but construct the station graph from a held-out portion of the records, or from physical distances alone, and then denoise the remaining data; if $F^\alpha_\otimes$ no longer concentrates roughly 99% of the energy in the first 40 bins or no longer beats the directed-graph fractional Fourier transform baseline, the central claim is undermined. Equally direct: compute $\|L^\alpha x\|$ for the clean correlated signal under the covariance-based weight; if it is not small, the truncation theorem's premise fails and the observed concentration must be attributed to the graph being matched to the signal.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the SVD-based graph Fourier transform idea, previously defined for directed product graphs, survives passage to fractional powers and splits into two viable variants. For factor graphs $G_1,G_2$ with Laplacians $L_1,L_2$, the paper forms fractional Laplacians $L_l^\alpha = P_l R_l Q_l^T$ by raising singular values to the power $\alpha$, and then either takes the SVD of the Kronecker-sum Laplacian $L^\alpha_\boxtimes = L^\alpha_1 \otimes I_{N_2} + I_{N_1}\otimes L^\alpha_2$ to get bases $P_\boxtimes,Q_\boxtimes$ (defining $F^\alpha_\boxtimes$), or combines factor SVDs through Kronecker products $P_1\otimes P_2$, $Q_1\otimes Q_2$ (defining $F^\alpha_\otimes$). Both transforms map $x$ to the pair $\frac12(P+Q)^T x$, $\frac12(P-Q)^T x$, so they are isometries and invertible. Theorems 3.2 and 4.2 bound the $\ell^2$ error of keeping only the first $\Omega$ frequencies by a constant times $(\|L^\alpha x\|+\|(L^\alpha)^T x\|)/r_{\Omega-1}$ (or the analogous $\tau_{\Omega-1}$ expression), which is the paper's formal evidence that strongly correlated signals are low-frequency dominated. In experiments the second transform has the lowest compute cost, the strongest energy concentration, and the best denoising numbers.

Load-bearing premise

The numerical demonstration assumes it is legitimate to build the graph from the same temperature signal that is subsequently noised and denoised, so the reported energy concentration and signal-to-noise gains may partly reflect matching the graph to the target data rather than an intrinsic property of the transform.

Editorial extensions

If this is right

  • When $\alpha=1$, both transforms reduce to the earlier SVD-based graph Fourier transforms $F_\boxtimes$ and $F_\otimes$, so the paper's results contain the non-fractional case and add a tunable fractional parameter.
  • The $\ell^2$ truncation bound gives a practical rule: a user can choose the bandwidth $\Omega$ from the cutoff frequency $r_{\Omega-1}$ and from measured norms like $\|L^\alpha_\boxtimes x\|$, with a guaranteed worst-case reconstruction error.
  • $F^\alpha_\otimes$ costs roughly $O(N_1^3+N_2^3)$ to build its basis instead of $O((N_1N_2)^3)$, so for large product graphs the separable construction is the scalable option.
  • The multi-graph generalization (Definitions 5.1 and 5.3) extends the same pair of constructions to Cartesian products of $m$ directed graphs, so tensors of many irregular dimensions can be handled by the same machinery.

Reading between the lines

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

  • A test the paper leaves unrun: construct the station graph from one part of the temperature record and denoise a disjoint part; if the SNR advantage over the baseline persists, the effect is tied to correlation structure, whereas the current setup builds weights from the very signal that is later noised and denoised.
  • The theorems establish low-frequency concentration from small fractional-Laplacian norms, but they do not prove that strong spatiotemporal correlation implies those norms are small; a stochastic model of correlated graph signals could close that gap.
  • The same energy concentration suggests lossy compression of multichannel sensor data: keeping the first $\Omega$ coefficients of $F^\alpha_\otimes$ would come with the paper's explicit worst-case error, a use the paper does not develop.
  • Because $F^\alpha_\otimes$ separates the two factor directions in Algorithm 1, one could compose different fractional orders along each axis, giving an anisotropic fractional spectral analysis for signals whose correlations differ across space and time.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes two graph fractional Fourier transforms, F^α_⊠ and F^α_⊗, for signals on Cartesian products of directed graphs, based on singular value decompositions of fractional Laplacian matrices. It proves inverse and Parseval-type identities, gives truncation bounds for low-frequency approximations (Theorems 3.2 and 4.2), extends the definitions to products of m directed graphs, and demonstrates denoising on hourly temperature data from 32 weather stations. The second transform F^α_⊗ is claimed to have lower computational complexity, and both are compared against the DGFRFT of Yan et al. and the GFTs of Cheng et al.

Significance. If the construction were valid, the paper would deliver a fractional spectral analysis tool for directed product graphs, with explicit truncation bounds and a computationally cheaper transform. Theorems 3.2 and 4.2 are useful spectral inequalities conditional on the orthonormality of the SVD factors, and the paper provides reproducible-looking definitions, algorithms, and real-data experiments. However, the central claim is currently not established for general directed graphs, because the fractional powers of the SVD factors need not be real orthogonal matrices. In addition, the experimental validation is partly circular because the graph weights and the fractional parameters are chosen using the clean target signal. The m-graph extension also lacks proofs.

major comments (4)
  1. [§3, Eq. (17); §4, Definition 4.1, Eqs. (39)-(40)] The definitions set P_l = U_l^α and Q_l = V_l^α, where U_l and V_l are the orthogonal factors from the SVD of a directed Laplacian. The proofs of invertibility (24), Parseval (25), and Theorems 3.2 and 4.2 require P_l and Q_l to be real orthonormal: for example, Eq. (29) uses ∑ r_i^2 (q_i^T x)^2 = ‖L^α_⊠ x‖^2, which assumes Q_⊠^T Q_⊠ = I. For a real orthogonal matrix with an eigenvalue −1, the principal matrix power has eigenvalue e^{iπα}, which is non-real for non-integer α; even in the absence of −1, the real-orthogonality of U^α depends on a branch choice that is not stated. No restriction on the SVD factors or on the matrix-power branch is given. Consequently F^α_⊠ and F^α_⊗ may not be real maps from R^N to R^{2N}, and the stated Parseval and invertibility identities do not follow for general directed graphs.
  2. [§6, Eqs. (59)-(60), Tables 3-6] The validation is partly circular. The weights w2 and w3 are computed from Cov(x^(i), x^(j)) and E(x^(i)) of the same clean signal X1 that is later corrupted and denoised, and the fractional order α = 0.7 and bandwidth Ω = 40 are fixed after inspecting the energy concentration in Figure 2. The theoretical results only bound the truncation error by terms involving ‖L^α x‖; they do not show that the temperature signal has small ‖L^α x‖ on a graph constructed from its own statistics. The reported SNR advantages in Tables 2-6 may therefore reflect leakage of the target signal into the graph construction rather than an intrinsic property of the proposed transforms.
  3. [§5, Theorems 5.2 and 5.4] The proofs of Theorems 5.2 and 5.4 are omitted with the statement that they are similar to Theorems 3.2 and 4.2. Since the m-graph extension is explicitly listed as a contribution, the proofs (or at least a complete proof of one theorem and a precise reduction for the other) should be supplied. These theorems also inherit the orthogonality issue of the first major comment, so an omitted proof cannot be assumed to go through.
  4. [Remark 2, Eq. (38)] The comparison baseline F^α_q is not fully specified. Remark 2 defines (P_{1,q} ⊗ P_{2,q})* x and asserts that it coincides with F^α_{⊠,q} and F^α_{⊗,q}, but those two transforms are never defined, and the claim that the tensor product of the two single-graph Hermitian DGFRFT bases is the eigenbasis of the Hermitian Laplacian of the product graph is not proved. This matters because Section 6 uses this baseline for the numerical comparison.
minor comments (5)
  1. [Throughout] There are several typographical errors: 'Cartrsian' in Definitions 2.1 and 2.2, 'Caresian' in Definitions 4.1 and 5.3, and 'In addtion' in the Conclusion.
  2. [§3-§4] The phrase 'strong spatiotemporal correlation' is not formalized; the theorems quantify energy concentration in terms of ‖L^α x‖, not in terms of a correlation model. The authors should either define a correlation model or rephrase the claim.
  3. [§6, Tables 3-5] The tables report averages over '100 times' and '31 days' but no standard deviations or confidence intervals, making it difficult to assess whether the SNR differences are statistically meaningful.
  4. [Theorem 3.2, Eq. (28)] The second inequality in (28) is obtained by applying the triangle inequality to the Kronecker sum decomposition of L^α_⊠, but this step is not stated; adding one sentence would improve readability.
  5. [Theorem 4.2, Eq. (43)] The inequality (‖a‖ + ‖b‖)^2 ≥ ∑ (r_{1,i} + r_{2,j})^2 (q^T x)^2 relies on the Cauchy-Schwarz inequality in a step that is not shown; expanding this step would make the proof easier to verify.

Circularity Check

1 steps flagged · score 4.0 of 10

Validation is partly circular: graph weights w2/w3 are fitted to the same temperature record used for denoising, so the reported energy concentration and SNR are partly forced; the transform definitions and SVD bounds themselves are self-contained.

  1. fitted input called prediction [Section 6, Eqs. (59)-(60), Figs. 2-3, Tables 2-5]
    "w2(i, j) = max( |Cov(x(i), x(j))| / (Var(x(i)) Var(x(j))) + u(i, j), 0 ), (59), and w3(i, j) = max( |E(x(i)) − E(x(j))| + u(i, j), 0 ), (60), where u(i,j) are i.i.d. with uniform distribution on [−0.2, 0.2]. ... by using the weight w2 (or w3), our numerical experiments indicate that the first 40 frequencies ... containing about 90.05%, 99.63%, and 81.27% (or 93.99%, 99.62%, and 81.29%) energy of x1."

    The weights w2/w3 use Cov(x(i),x(j)) and |E(x(i))−E(x(j))| of the temperature record x(i), which includes the very day X1 later noised and denoised. These weights define the graph Laplacians whose SVD factors P_l,Q_l generate the GFRFT bases in (17)/(33). The reported 99.62% energy concentration in the first 40 F^alpha_otimes frequencies under w2/w3 is therefore a consequence of fitting the graph to X1's covariance/mean, not an independent property of the transform. Tables 2-5 similarly denoise bX1=X1+noise using bases built from clean X1, so high SNR is partly forced. Theorems 3.2/4.2 only bound error by ||L^alpha x||; they never show strong correlation implies small seminorm, so they do not repair the circular validation.

full rationale

The mathematical core of the paper is not circular. Definitions 3.1 and 4.1 are explicit fractional extensions of Cheng et al.'s SVD-based GFTs, and Theorems 3.2, 4.2, 5.2, and 5.4 are standard SVD truncation inequalities: they prove that a signal is well approximated by its first Omega components whenever the relevant L^alpha seminorms are small. No load-bearing self-citation or imported uniqueness theorem appears; Yan et al. and Cheng et al. are independent prior works. The circularity is confined to Section 6. The covariance/mean-based weights w2 and w3 (Eqs. 59-60) are fitted to the same temperature record that is subsequently noised and bandlimited, so the headline energy-concentration percentages and denoising SNRs under those weights are partly artifacts of the graph-construction fit. In addition, the paper's conclusion states that strong spatiotemporal correlation guarantees stable recovery, but the theorems only establish a conditional bound in terms of ||L^alpha x||; the missing implication from correlation to small seminorm is never proved. Because the transform definitions, the alpha=1 reduction, and the w1-based experiments provide independent content, the overall circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The transforms themselves rest on standard matrix algebra and on the paper's SVD-based convention for fractional Laplacians. The free parameters are concentrated in the numerical validation: alpha, bandwidth Omega, and the data-adaptive graph weights. No new physical entities are postulated; the novel mathematical objects are definitional choices, listed as ad hoc axioms.

free parameters (3)
  • fractional order alpha = 0.7 (values 0.2 to 1 explored in Table 6)
    The GFRFTs are parameterized by alpha in (0,1]; experiments fix alpha=0.7 for all reported denoising, which affects energy concentration and SNR. No selection criterion is given, and the improvement over alpha=1 is not uniform across weights.
  • frequency bandwidth Omega = 40 (sweep 28-64)
    Denoising keeps the first Omega singular components; Omega=40 is the main setting in Tables 2, 3, 6 and Figure 3, chosen after observing that roughly 99% of energy lies in the first 40 coefficients.
  • station graph weights w2 and w3 = data-dependent functions of the clean signal
    Equation (59) uses Cov and Var of the clean signal x(i), and Eq. (60) uses expected values; these are not independent model constants but functions of the target data, making the graph adapt to the signal being denoised.
assumptions (3)
  • standard math Existence and properties of SVD and Kronecker sums for Laplacians L^alpha_l and L^alpha_boxtimes.
    Used throughout Sections 3-5 to define the matrices P, Q, R and to prove Theorems 3.2, 4.2, 5.2, and 5.4.
  • ad hoc to paper The SVD-based fractional Laplacian L^alpha_l := U_l Sigma_l^alpha V_l^T is a valid fractional graph operator for directed graphs.
    Equation (17) defines the fractional Laplacian by applying the power alpha to singular values only; this is the paper's convention, not the standard matrix power, and no independent justification is offered.
  • domain assumption Smallness of ||L^alpha x|| and ||(L^alpha)^T x|| is a faithful model of strong spatiotemporal correlation.
    Theorems 3.2 and 4.2 only bound truncation error in terms of these norms; the paper equates strong correlation with this graph smoothness but does not prove that meteorological temperature signals satisfy it under the constructed weights.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SVD-Based Graph Fractional Fourier Transform on Directed Graphs and Its Application." pith.science (2026). https://pith.science/paper/JKMYS3RI

@misc{pith2026250603925,
  author       = {Pith},
  title        = {Pith review of: SVD-Based Graph Fractional Fourier Transform on Directed Graphs and Its Application},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKMYS3RI}},
  note         = {Machine review of arXiv:2506.03925}
}
read the original abstract

Real-world signals frequently reside on directed Cartesian product graphs, including digital images, sensor networks, and meteorological temperature records. Designing a transform method suitable for processing such multi-dimensional graph signals within the fractional Fourier transform domain remains a critical challenge in graph signal processing (GSP). This paper proposes two novel graph fractional Fourier transforms (GFRFTs) for multi-dimensional signals defined on such directed product graphs and comprehensively investigates their denoising capabilities. Our contributions are fourfold: (1) We propose two distinct two-dimensional GFRFTs based on singular value decompositions of some fractional Laplacian matrices; (2) We generalize these transforms to multi-dimensional graph fractional Fourier transforms (MGFRFTs), establishing a powerful fractional domain analysis framework for multi-dimensional GSP; (3) We investigate the signal reconstruction capability of our proposed GFRFTs, as well as their computational complexity; and (4) We validate the practical utility of our approach through denoising experiments on real-world meteorological temperature datasets.

Figures

Figures reproduced from arXiv: 2506.03925 by the authors.

Figure 1
Figure 1. Original and noisy temperature signals collected at 32 weather stations in the [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. On the top are the first component 1 2 (P⊠ + Q⊠) T x1 and the second component 1 2 (P⊠ − Q⊠) T x1 of the GFRFT F α ⊠x1 (22). On the middle are the first component 1 2 (P⊗ + Q⊗) T x1 and second component 1 2 (P⊗ − Q⊗) T x1 of the GFRFT F α ⊗x1 (35). On the bottom are the real part and imaginary part of (P1,q ⊗ P2,q) ∗x1 for the DGFRFT F α q x1 (38). 22 [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Denoised signals Xe α 1,Ω,⊠, Xe α 1,Ω,⊗ and Xe α 1,Ω,q with weights w1 and w2. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    Graph signal processing: Overview, challenges, and applications,

    A. Ortega, P. Frossard, J. Kovaˇ cevi´ c, et al, “Graph signal processing: Overview, challenges, and applications,” Proc. IEEE, vol. 106, no. 5, pp. 808-828, May 2018

  2. [2]

    Graph signal processing in applications to sensor networks, smart grids, and smart cities,

    I. Jab lo´ nski, “Graph signal processing in applications to sensor networks, smart grids, and smart cities,” in IEEE Sens. J. , vol. 17, no. 23, pp. 7659-7666, Dec. 2017

  3. [3]

    Graph signal processing for machine learning: A review and new perspectives,

    X. Dong, D. Thanou, L. Toni, et al, “Graph signal processing for machine learning: A review and new perspectives,” IEEE Signal Process. Mag. , vol. 37, no. 6, pp. 117-127, Nov. 2020

  4. [4]

    A graph signal processing perspective on functional brain imaging,

    W. Huang, T. A. W. Bolton, J. D. Medaglia, et al, “A graph signal processing perspective on functional brain imaging,” Proc. IEEE, vol. 106, no. 5, pp. 868-885, May 2018

  5. [5]

    Grid-graph signal processing (Grid- GSP): A graph signal processing framework for the power grid,

    R. Ramakrishna and A. Scaglione, “Grid-graph signal processing (Grid- GSP): A graph signal processing framework for the power grid,” IEEE Trans. Signal Process., vol. 69, pp. 2725-2739, 2021. 26

  6. [6]

    Graph Fourier transform based on ℓ1 norm variation minimization,

    L. Yang, A. Qi, C. Huang, et al, “Graph Fourier transform based on ℓ1 norm variation minimization,” Appl. Comput. Harmon. Anal. , vol. 52, pp. 348-365, 2021

  7. [7]

    Multi-dimensional visual data completion via weighted hybrid graph-Laplacian,

    J. Jiang, Y. Xu, H. Xu, et al, “Multi-dimensional visual data completion via weighted hybrid graph-Laplacian,” Signal Process., vol. 216, 2024, Art. no. 109305

  8. [8]

    Big data analysis with signal pro- cessing on graphs: Representation and processing of massive data sets with irregular structure,

    A. Sandryhaila and J. M. F. Moura, “Big data analysis with signal pro- cessing on graphs: Representation and processing of massive data sets with irregular structure,” IEEE Signal Process. Mag. , vol. 31, no. 5, pp. 80-90, Sep. 2014

Show all 26 references
  1. [9]

    Harmonic analysis on directed graphs and applications: From Fourier analysis to wavelets,

    H. Sevi, G. Rilling, and P. Borgnat, “Harmonic analysis on directed graphs and applications: From Fourier analysis to wavelets,”Appl. Com- put. Harmon. Anal. , vol. 62, pp. 390-440, 2023

  2. [10]

    A directed graph Fourier transform with spread frequency components,

    R. Shafipour, A. Khodabakhsh, G. Mateos, et al, “A directed graph Fourier transform with spread frequency components,”IEEE Trans. Sig- nal Process., vol. 67, no. 4, pp. 946-960, Feb. 2019

  3. [11]

    Representation of functions on big data associated with directed graphs,

    C. K. Chui, H. Mhaskar, and X. Zhuang, “Representation of functions on big data associated with directed graphs,” Appl. Comput. Harmon. Anal., vol. 44, no. 1, pp. 165-188, 2018

  4. [12]

    Signal processing on directed graphs: the role of edge directionality when processing and learning from network data,

    A. Marques, S. Segarra, and G. Mateos, “Signal processing on directed graphs: the role of edge directionality when processing and learning from network data,” IEEE Signal Process. Mag. , vol. 37, no. 6, pp. 99–116, 2020

  5. [13]

    On the graph Fourier transform for directed graphs,

    S. Sardellitti, S. Barbarossa, and P. DiLorenzo, “On the graph Fourier transform for directed graphs,” IEEE J. Sel. Top. Signal Process. , vol. 11, no. 6, pp. 796–811, 2017

  6. [14]

    Graph Fourier transform based on singular value decomposition of directed Laplacian,

    Y. Chen, C. Cheng, and Q. Sun, “Graph Fourier transform based on singular value decomposition of directed Laplacian,” Sampling Theory Signal Process. Data Anal. , vol. 12, 2023, Art. no. 24

  7. [15]

    Spectral graph fractional Fourier transform for di- rected graphs and its application,

    F. Yan and B. Li, “Spectral graph fractional Fourier transform for di- rected graphs and its application,” Signal Process., vol. 210, 2023, Art. no. 109099. 27

  8. [16]

    The optimal joint time-vertex graph filter design: From ordinary graph Fourier domains to fractional graph Fourier domains,

    Z. Ge, H. Guo, T. Wang, et al, “The optimal joint time-vertex graph filter design: From ordinary graph Fourier domains to fractional graph Fourier domains,” Circu. Syst. Signal Process. , vol. 42, no. 7, pp. 4002- 4018, 2023

  9. [17]

    Multi-dimensional graph fractional Fourier trans- form and its application,

    F. Yan and B. Li, “Multi-dimensional graph fractional Fourier trans- form and its application,” Digit. Signal Process., vol. 129, 2022, Art. no. 103683

  10. [18]

    Generalized sampling of graph signals with the prior information based on graph fractional Fourier transform,

    D. Wei and Z. Yan, “Generalized sampling of graph signals with the prior information based on graph fractional Fourier transform,” Signal Process., vol. 214, 2024, Art. no. 109263

  11. [19]

    Graph fractional Fourier trans- form: A unified theory,

    T. Alika¸ sifoˇ glu, B. Kartal, and A. Ko¸ c, “Graph fractional Fourier trans- form: A unified theory,” IEEE Trans. Signal Process., vol. 72, pp. 3834- 3850, 2024

  12. [20]

    Local measurement and reconstruction for noisy bandlimited graph signals,

    X. Wang, J. Chen, and Y. Gu, “Local measurement and reconstruction for noisy bandlimited graph signals,” Signal Process., vol. 129, pp. 119- 129, 2016

  13. [21]

    Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,

    A. Anis, A. Gadde, and A. Ortega, “Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,” IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3775-3789, 2016

  14. [22]

    Time-varying graph signal reconstruc- tion,

    K. Qiu, X. Mao, X. Shen, et al, “Time-varying graph signal reconstruc- tion,” IEEE J. Sel. Top. Signal Process., vol. 11, no. 6, pp. 870-883, Sep. 2017

  15. [23]

    Time-varying graph signal denoising via me- dian filters,

    D. B. Tay and J. Jiang, “Time-varying graph signal denoising via me- dian filters,” IEEE Trans. Circu. Syst. II: Express Briefs , vol. 68, no. 3, pp. 1053-1057, Mar. 2021

  16. [24]

    SVD-based graph Fourier trans- forms on directed product graphs,

    C. Cheng, Y. Chen, Y. J. Lee, et al, “SVD-based graph Fourier trans- forms on directed product graphs,” IEEE Trans. Signal Inf. Process. Networks, vol. 9, pp. 531-541, 2023

  17. [25]

    Stationary signal processing on graphs,

    N. Perraudin and P. Vandergheynst, “Stationary signal processing on graphs,” IEEE Trans. Signal Process. , vol. 65, no. 13, pp. 3462-3477, Jul. 2017

  18. [26]

    Graph signal processing for directed graphs based on the Hermitian Laplacian,

    S. Furutani, T. Shibahara, M. Akiyama, et al, “Graph signal processing for directed graphs based on the Hermitian Laplacian,” in Proc. Joint Eur. Conf. Mach. Learn. Knowl. Discov. Databases , 2020, pp. 447-463. 28 Table 5: The average bandlimiting SNR for different weights wi,...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.