betaFeedback

WorksMathematics

Computing Data-Aware Diagonal Frame Decompositions

Abstract

We present a method for computing a specific type of matrix decomposition that combines a structured framework with data-driven adaptation. This framework allows for an efficient and stable representation of signals or images using frames by identifying sets of vectors that naturally capture the essential features of the given data. The resulting decomposition enables the development of regularization techniques that enhance the quality of reconstruction by preserving important signal characteristics while suppressing noise. We develop an optimization-based algorithm to compute the decomposition and establish convergence properties of the resulting iterates. Numerical experiments on synthetic data and tomographic image reconstruction illustrate the behavior of the proposed method and its performance compared to conventional spectral regularization methods.

Introduction

In this paper, we consider a matrix decomposition problem that we call diagonal tight frame decomposition [27] with data-aware adaptation for sparse representation. The decomposition is related to the singular value decomposition, with the difference that the corresponding singular vectors form a tight frame (for a precise definition, see § 2.1) instead of an orthonormal basis. The problem is formulated as follows:

Problem 1. Given a matrix A∈RN×M\mathbf{A} \in\mathbb{R}^{N \times M} and a data set X=[x1,x2,…,xS]∈RM×S\mathbf{X} = [\mathbf{x}_1,\mathbf{x}_2,\ldots,\mathbf{x}_S] \in\mathbb{R}^{M \times S}, find matrices U\mathbf{U}, D\mathbf{D}, and V\mathbf{V} satisfying the following properties:

  1. Tightness: U\mathbf{U} and V\mathbf{V} are tight frames,

  1. Scalability: D\mathbf{D} is diagonal,

  1. Decomposition: ATV=UD\mathbf{A}^{\mathsf{T}}\mathbf{V} = \mathbf{U}\mathbf{D},

  1. Sparsity: the representations UTX\mathbf{U}^{\mathsf{T}}\mathbf{X} are sparse.

The motivation for this problem comes from the use of frame decompositions for the representation and inversion of linear operators in Hilbert spaces. This diagonalized structure allows the evaluation and inversion of the operator to be analysed and implemented in terms of frame operators and scalar multiplication. The sparsity property incorporates prior information about the underlying signal class and can help stabilize the representation. Such decompositions have been used, for example, in the regularization of inverse problems [9, 20, 21, 22, 26, 27].

If U and V are orthonormal bases for RM\mathbb{R}^{M} and RN\mathbb{R}^{N}, respectively, then A=VDUT\mathbf{A} = \mathbf{V}\mathbf{D}\mathbf{U}^{\mathsf{T}} is a singular value decomposition. However, the singular vectors generally do not provide sparse representations for a given signal class. They are determined solely by the matrix A\mathbf{A} and therefore cannot be adapted to different signal distributions. In contrast, tight frames provide stable and redundant representations that are well suited for sparse signal representations. Moreover, their dual frames are particularly easy to compute, since the canonical dual of a tight frame differs only by a multiplicative constant. The diagonal matrix D\mathbf{D} plays a role analogous to the singular values and is therefore often referred to as the matrix of quasi-singular values. It represents the scaling action of the operator in the corresponding frame coordinates. Replacing orthonormal bases by tight frames therefore provides greater flexibility in the construction of such decompositions, allowing additional properties to be imposed. In particular, our sparsity condition introduces a data-dependent criterion: rather than constructing a decomposition solely from the algebraic properties of A\mathbf{A}, we seek a frame U\mathbf{U} whose coefficients provide sparse representations given the data X\mathbf{X}. This is motivated by the observation that suitable representations can capture the structure of signals and images more efficiently and incorporate prior information that is relevant for inverse problems [12, 35]. We note in passing that the diagonal matrix decomposition problems has applications far beyond imaging. In fact, frames are widely used for this purpose in applied mathematics, computer science, and engineering, with applications including sampling theory [23], data quantization [5, 6], image processing [10, 29], and coding theory [40].

Applications

Application 1: Inverse problems. In inverse problems, one seeks to recover an unknown parameter x∈Xx \in X from noisy measurements

Kx+n=yδ.(1)\mathbf{K}x+n=y^{\delta}. \tag*{(1)}

K:X→Y\mathbf{K}: X \to Y is a linear operator between Hilbert spaces XX and YY, nn denotes a bounded noise term, and yδ∈Yy^{\delta} \in Y is the observed data. The goal is to recover the unknown xx from the noisy measurements yδy^{\delta}.

The operator K\mathbf{K} is typically not continuously invertible, making the reconstruction problem ill-posed and necessitating regularization. Given a discretization A∈RN×M\mathbf{A} \in\mathbb{R}^{N \times M} of K\mathbf{K}, we follow the discretize-then-optimize methodology from numerical analysis. Since A\mathbf{A} is typically ill-conditioned, direct inversion is highly sensitive to measurement noise and therefore does not provide a stable reconstruction.

Given the decomposition (U,V,D)(\mathbf{U},\mathbf{V},\mathbf{D}) searched for in Problem 1, we have ATV=UD\mathbf{A}^{\mathsf{T}}\mathbf{V}=\mathbf{U}\mathbf{D} for tight frames U\mathbf{U} and V\mathbf{V} and diagonal matrix D\mathbf{D}. Let x+∈ker⁡(A)⊥x^{+} \in\operatorname{ker}(\mathbf{A})^{\perp} be the least-squares solution of Ax=y\mathbf{A}x=y for noiseless observations y∈ran⁡(A)y \in\operatorname{ran}(\mathbf{A}). By the sparsity constraint, the frame coefficients of x+x^{+} are assumed to be sparse with respect to U\mathbf{U}. In particular, given a dual frame Uˉ\bar{\mathbf{U}} of U\mathbf{U}, the duality relation for frames gives

x+=UˉUˉTx+=∑i=1K⟨x+,ui⟩uˉi,x^{+}=\bar{\mathbf{U}}\bar{\mathbf{U}}^{\mathsf{T}}x^{+}=\sum_{i=1}^{K}\langle x^{+},u_i\rangle\bar{u}_i,

where uiu_i and uˉi\bar{u}_i denote the columns of U\mathbf{U} and Uˉ\bar{\mathbf{U}}, respectively. Thus, if the frame coefficients ⟨x+,ui⟩\langle x^+,u_i\rangle are sparse, the signal can be represented accurately using only a small number of frame elements. The decomposition further yields the relation D†VTA=UT\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}\mathbf{A}=\mathbf{U}^{\mathsf{T}}, where D†\mathbf{D}^{\dagger} denotes the Moore–Penrose inverse of D\mathbf{D}. Consequently,

UTx+=D†VTAx+=D†VTy,\mathbf{U}^{\mathsf{T}}x^+=\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}\mathbf{A}x^+=\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}y,

and hence D†VTy\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}y is sparse as well. The Moore–Penrose inverse of A\mathbf{A} is given by A+=UˉD†VT\mathbf{A}^{+}=\bar{\mathbf{U}}\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}. For noisy data yδ=Ax++ny^\delta=\mathbf{A}x^++n, the corresponding reconstruction is

A+yδ=UˉD†VTyδ=Uˉ(D†VTy+D†VTn).\mathbf{A}^{+}y^\delta=\bar{\mathbf{U}}\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}y^\delta=\bar{\mathbf{U}}\left(\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}y+\mathbf{D}^{\dagger}\mathbf{V}^{\mathsf{T}}n\right).

This representation separates the reconstruction into the frame coefficients of the underlying signal and a noise term amplified by D†\mathbf{D}^{\dagger}. Small entries of D\mathbf{D} reflect the ill-conditioning of A\mathbf{A} and can lead to strong amplification of the noise. The sparsity of the signal representation therefore provides a natural basis for regularization: rather than applying the inverse directly, the coefficients can be filtered to suppress noise while preserving the significant components of the signal. Such regularized reconstructions take the form

Rα(yδ)=Uˉfα(D†,VTyδ),(2)\mathcal{R}_{\alpha}(y^\delta)=\bar{\mathbf{U}}f_{\alpha}\left(\mathbf{D}^{\dagger},\mathbf{V}^{\mathsf{T}}y^\delta\right), \tag*{(2)}

where the filter is applied component-wise. An example is the soft-thresholding filter

fα(d,c)=sign⁡(c)max⁡{∣c∣−α/d,0}.(3)f_{\alpha}(d,c)=\operatorname{sign}(c)\max\{|c|-\alpha/d,0\}. \tag*{(3)}

Under suitable assumptions on the filter and the parameter choice α=α(δ)\alpha=\alpha(\delta), such filtered reconstructions can yield a convergent regularization method in the sense of [24] as the noise level δ\delta tends to zero, see [22, 27].

Note that, although frame decompositions with non-tight frames are also used for inverse problems, non-tight frames in finite dimensions have the disadvantage that the frame bounds must be controlled and the corresponding dual frame must be computed.

Application 2: Sparse representation of observations. The matrix decomposition also provides a sparse representation of the observations. Indeed, from ATV=UD\mathbf{A}^{\mathsf{T}}\mathbf{V}=\mathbf{U}\mathbf{D} we obtain A=VˉDUT\mathbf{A}=\bar{\mathbf{V}}\mathbf{D}\mathbf{U}^{\mathsf{T}}, and hence

y=Ax=VˉDUTx.y=\mathbf{A}x=\bar{\mathbf{V}}\mathbf{D}\mathbf{U}^{\mathsf{T}}x.

If the frame coefficients UTx\mathbf{U}^{\mathsf{T}}x are sparse, then the coefficient vector DUTx=VTy\mathbf{D}\mathbf{U}^{\mathsf{T}}x=\mathbf{V}^{\mathsf{T}}y is sparse as well. Thus, the observations yy admit a sparse representation with respect to the dual frame Vˉ\bar{\mathbf{V}},

y=∑i=1Kdi⟨x,ui⟩vˉi,y=\sum_{i=1}^{K}d_i\langle x,u_i\rangle\bar{v}_i,

where vˉi\bar{v}_i denotes the iith element of the dual frame Vˉ\bar{\mathbf{V}}. Consequently, the observations can be represented and stored using only a small number of nonzero frame coefficients. In particular, instead of storing the full observation vector yy, one may store its sparse frame representation VTy=DUTx\mathbf{V}^{\mathsf{T}}y=\mathbf{D}\mathbf{U}^{\mathsf{T}}x together with the corresponding indices.

Main contributions

In this paper, we develop a numerical methodology for solving the structured matrix decomposition (Problem 1). Our approach combines an algebraic reduction of the decomposition problem with a new non-smooth and non-convex hierarchical (bilevel) optimization formulation of the original matrix decomposition problem. In particular, we develop from scratch a new bilevel optimization formulation on the Stiefel manifold containing the desired frame matrices.

We first show that Problem 1 can be reduced to an equivalent problem involving only positive diagonal matrices for A. This reduction allows us to characterize necessary and sufficient conditions on the diagonal matrix D for the existence of a decomposition satisfying the prescribed tightness and decomposition constraints. Once D is fixed, the remaining problem can be formulated entirely in terms of the frame U, since the second frame V is uniquely determined by U and D. The tightness constraint on U naturally leads to an optimization problem on the Stiefel manifold.

To incorporate the sparsity requirement, we formulate the resulting problem as a simple bilevel optimization problem on the Stiefel manifold (11). The lower-level problem is a smooth but non-convex optimization problem whose solution guarantees the tightness of V and the decomposition property, while the upper-level objective is convex and lower semi-continuous, incorporating the sparsity requirement. The tightness of U follows directly from the definition of the Stiefel manifold. Although the lower-level problem is convex when considered in the ambient Euclidean space, this convexity is generally lost after restricting the problem to the Stiefel manifold. We address the resulting nonconvex bilevel problem using a penalized subgradient method formulated in a general setting, which therefore applies beyond the specific matrix decomposition considered here.

We provide a rigorous convergence analysis of both levels of the resulting bilevel scheme. In particular, we establish convergence of the lower-level optimization procedure and analyze the convergence of the iterates to stationary points of the bilevel problem under the stated assumptions. The theoretical results are complemented by numerical experiments that demonstrate the practical behavior of the proposed method and its ability to identify sparse frame representations adapted to the given data.

Related work

Optimization-based approaches to frame computation have been studied, for example, by [41], who develop convex optimization methods for constructing frames with prescribed structural properties. Our setting differs in that the frame is not designed independently of an operator; instead, the frame and a diagonal scaling are jointly required to represent a given matrix while promoting sparsity of its action on a data set. A related line of work concerns scalable frames, where one seeks diagonal scalings that transform a given frame into a Parseval frame; see, e.g., [30]. Scalable-frame theory typically considers a fixed frame and asks whether a positive diagonal scaling yields a Parseval frame. In contrast, we optimize the frame itself together with the diagonal scaling, with the additional requirement that their product represent a prescribed operator and with a data-dependent sparsity objective.

Optimization subject to orthogonality constraints is naturally formulated on matrix manifolds. Structural optimization on embedded submanifolds was studied, for example, in [8]. For optimization on the Stiefel manifold, proximal gradient methods have been developed in [13]; see [14] for a broader treatment of proximal and related methods for nonsmooth optimization on matrix manifolds. Weakly convex optimization, which is relevant to the nonsmooth structure arising in our analysis, has been studied extensively in [18, 31].

Penalty methods are classical approaches to solve constrained optimization problems via a sequence of unconstrained problems. In convex optimization in Euclidean spaces, prox-penalization and related diagonal-gradient approaches have been studied extensively [1, 2, 3, 4, 34, 38]. In contrast, penalty formulations specifically tailored to optimization on the Stiefel manifold are comparatively less developed. For example, a Lagrangian relaxation approach for optimization with Stiefel constraints is considered in [43]. Our work combines these ideas by employing a penalty formulation together with a subgradient method on the Stiefel manifold to solve the bilevel frame decomposition problem.

Preliminaries

Throughout, we work in real finite-dimensional vector spaces equipped with an inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle and the associated norm ∥⋅∥\|\cdot\|. Unless stated otherwise, we use the Frobenius norm for matrices and the ℓ2\ell^2-norm for vectors, i.e.,

∥A∥F:=(∑i,j∣Aij∣2)1/2,∥x∥2:=(∑i∣xi∣2)1/2,\|A\|_{\mathrm{F}} := \left(\sum_{i,j}|A_{ij}|^2\right)^{1/2}, \qquad\|x\|_2 := \left(\sum_i |x_i|^2\right)^{1/2},

where we omit the subscript when it is clear from the context.

Frames

We briefly recall some basic facts about finite-dimensional frame theory. More background on frame theory in Hilbert spaces can be found in [15], while [42] provides an introduction to finite tight frames. Let R,K∈NR,K\in\mathbb{N} with K≥RK\ge R. A family (ui)i=1K⊂RR(u_i)_{i=1}^{K}\subset\mathbb{R}^{R} is called a frame for RR\mathbb{R}^{R} if there exist constants A,B>0A,B>0 such that

A∥x∥2≤∑i=1K∣⟨ui,x⟩∣2≤B∥x∥2for all x∈RR.(4)A\|x\|^2 \le\sum_{i=1}^{K}|\langle u_i,x\rangle|^2 \le B\|x\|^2 \qquad\text{for all }x\in\mathbb{R}^{R}. \tag*{(4)}

In finite dimensions, we identify the frame with the matrix U∈RR×KU\in\mathbb{R}^{R\times K} whose columns are the frame elements (ui)i=1K(u_i)_{i=1}^{K}. The constants AA and BB are called the lower and upper frame bounds of UU, respectively. The frame is called tight if A=BA=B. A simple example of a tight frame is the union of pp orthonormal bases, in which case the frame bounds are A=B=pA=B=p. Other prominent examples include finite-dimensional versions of wavelet, curvelet, and Gabor frames, which arise, for example, from suitable discretizations of these constructions in imaging applications [9, 10, 17]. A tight frame with A=B=1A=B=1 is called a Parseval frame. Furthermore, a frame is called exact if deleting any single frame element results in a family that is no longer a frame. A frame that is not a basis is called overcomplete.

For a frame represented by U∈RR×KU\in\mathbb{R}^{R\times K}, the corresponding analysis and synthesis operations are given by the matrix products

UTx∈RKandUc∈RR,U^{\mathsf{T}}x\in\mathbb{R}^{K} \qquad\text{and}\qquad Uc\in\mathbb{R}^{R},

respectively. Thus, UTU^{\mathsf{T}} maps a vector to its frame coefficients, while UU synthesizes a vector from a given coefficient vector. The associated frame operator is therefore

S=UUT,S=UU^{\mathsf{T}},

which is symmetric positive definite, and hence invertible. In particular, the frame condition can equivalently be written as AIdRR⪯UUT⪯BIdRRA\mathrm{Id}_{\mathbb{R}^{R}} \preceq UU^{\mathsf{T}} \preceq B\mathrm{Id}_{\mathbb{R}^{R}}. For a tight frame with frame constant AA, the frame condition is equivalent to UUT=AIdRRUU^{\mathsf{T}} = A\mathrm{Id}_{\mathbb{R}^{R}}, while for a Parseval frame it is UUT=IdRRUU^{\mathsf{T}} = \mathrm{Id}_{\mathbb{R}^{R}}. Notably, every tight frame can be scaled to a Parseval frame: if UU is a tight frame with frame constant AA, then U/AU/\sqrt{A} is a Parseval frame.

The redundancy of a frame generally leads to multiple possible coefficient representations of a given vector. In particular, there are different frames U‾∈RR×K\overline{U} \in\mathbb{R}^{R \times K} that can serve as dual frames and satisfy the reconstruction identity

x=U‾UTxfor all x∈RR,x = \overline{U}U^{\mathsf{T}}x \quad\text{for all } x \in\mathbb{R}^{R},

References

  1. [1]Hedy Attouch. Viscosity solutions of minimization problems. SIAM Journal on Optimization, 6(3):769–806, 1996.DOI
  2. [2]Hedy Attouch, Alexandre Cabot, and Marc-Olivier Czarnecki. Asymptotic behavior of nonautonomous monotone and subgradient evolution equations. Transactions of the American Mathematical Society, 370(2):755–790, 2018.arxiv.org/abs/1601.00767
  3. [3]Hédy Attouch, Marc-Olivier Czarnecki, and Juan Peypouquet. Coupling forward-backward with penalty schemes and parallel splitting for constrained variational inequalities. SIAM Journal on Optimization, 21(4):1251–1274, 2011.DOI
  4. [4]Jean-Bernard Baillon and Roberto Cominetti. A convergence result for nonautonomous subgradient evolution equations and its application to the steepest descent exponential penalty trajectory in linear programming. Journal of Functional Analysis, 187(2):263–273, 2001.DOI
  5. [5]James Blum, Mark Lammers, Alexander M Powell, and Özgür Yılmaz. Sobolev duals in frame theory and sigma-delta quantization. Journal of Fourier Analysis and Applications, 16(3):365–381, 2010.DOI
  6. [6]Bernhard G Bodmann and Vern I Paulsen. Frame paths and error bounds for sigma–delta quantization. Applied and Computational Harmonic Analysis, 22(2):176–197, 2007.DOI
  7. [7]Nicolas Boumal. An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023.DOI
  8. [8]Léopold Cambier and P.-A. Absil. Robust low-rank matrix completion by riemannian optimization. SIAM Journal on Scientific Computing, 38(5):S440–S460, 2016.DOI
  9. [9]Emmanuel J Candès and David L Donoho. Recovering edges in ill-posed inverse problems: Optimality of curvelet frames. The Annals of Statistics, 30(3):784–842, 2002.DOI
  10. [10]Emmanuel J Candès and David L Donoho. New tight frames of curvelets and optimal representations of objects with piecewise c2 singularities. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 57(2):219–266, 2004.
  11. [11]Jincheng Cao, Ruichen Jiang, Erfan Yazdanoust Hamedani, and Aryan Mokhtari. On the complexity of finding stationary points in nonconvex simple bilevel optimization. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 32126–32153. Curran Associates, Inc., 2025.arxiv.org/abs/2507.23155
  12. [12]C. Chaux, L. Duval, and J.-C. Pesquet. Image analysis using a dual-tree m-band wavelet transform. IEEE Transactions on Image Processing, 15(8):2397–2412, 2006.
  13. [13]Shixiang Chen, Shiqian Ma, Anthony Man-Cho So, and Tong Zhang. Proximal gradient method for nonsmooth optimization over the stiefel manifold. SIAM Journal on Optimization, 30(1):210–239, 2020.arxiv.org/abs/1811.00980
  14. [14]Shixiang Chen, Shiqian Ma, Anthony Man-Cho So, and Tong Zhang. Nonsmooth optimization over the stiefel manifold and beyond: Proximal gradient method and recent variants. SIAM Review, 66(2):319–352, 2024.DOI
  15. [15]O. Christensen. An Introduction to Frames and Riesz Bases. Birkhäuser, 2002.
  16. [16]Daniel Cortild, Mathias Staudigl, Juan Peypouquet, and Coralia Cartis. Stochastic nonconvex bilevel optimization: Improved rates without rare-visit assumption, 2026.arxiv.org/abs/2609.06580
  17. [17]Ingrid Daubechies. Ten lectures on wavelets. SIAM, 1992.
  18. [18]Damek Davis and Dmitriy Drusvyatskiy. Stochastic model-based minimization of weakly convex functions. SIAM Journal on Optimization, 29(1):207–239, 2019.arxiv.org/abs/1803.06523
  19. [19]Stephan Dempe, Nguyen Dinh, Joydeep Dutta, and Tanushree Pandit. Simple bilevel programming and extensions. Mathematical Programming, 188(1):227–253, 2021.
  20. [20]David L Donoho. Nonlinear solution of linear inverse problems by wavelet–vaguelette decomposition. Applied and computational harmonic analysis, 2(2):101–126, 1995.DOI
  21. [21]Andrea Ebner, Jürgen Frikel, Dirk Lorenz, Johannes Schwab, and Markus Haltmeier. Regularization of inverse problems by filtered diagonal frame decomposition. Applied and Computational Harmonic Analysis, 62:66–83, 2023.arxiv.org/abs/2008.06219
  22. [22]Andrea Ebner and Markus Haltmeier. Convergence of non-linear diagonal frame filtering for regularizing inverse problems. Inverse Problems, 40(5), 2024.arxiv.org/abs/2308.15666
  23. [23]Yonina C Eldar. Sampling with arbitrary sampling and reconstruction spaces and oblique dual frame vectors. Journal of Fourier Analysis and Applications, 9(1):77–96, 2003.DOI
  24. [24]Heinz W. Engl, Martin Hanke, and Andreas Neubauer. Regularization of Inverse Problems. Springer, 1996.
  25. [25]Ky Fan and Gordon Pall. Imbedding conditions for hermitian and normal matrices. Canadian Journal of Mathematics, 9:298–304, 1957.DOI
  26. [26]Jürgen Frikel and Markus Haltmeier. Efficient regularization with wavelet sparsity constraints in photoacoustic tomography. Inverse Problems, 34(2), 2018.DOI
  27. [27]Jürgen Frikel and Markus Haltmeier. Sparse regularization of inverse problems by operator-adapted frame thresholding. Trends in Mathematics, pages 163–178, 2020.arxiv.org/abs/1909.09364
  28. [28]Peter Kuchment. The Radon Transform and Medical Imaging. SIAM, 2013.DOI
  29. [29]Gitta Kutyniok and Wang-Q Lim. Compactly supported shearlets are optimally sparse. Journal of Approximation Theory, 163(11):1564–1589, 2011.arxiv.org/abs/1002.2661
  30. [30]Gitta Kutyniok, Kasso A. Okoudjou, Friedrich Philipp, and Elizabeth K. Tuley. Scalable frames. Linear Algebra and its Applications, 438(5):2225–2238, 2013.
  31. [31]Xiaoli Li, Shixiang Chen, Zengde Deng, Qing Qu, Zhihui Zhu, and Anthony Man-Cho So. Weakly convex optimization over stiefel manifold using riemannian subgradient-type methods. SIAM Journal on Optimization, 31(3):1605–1634, 2021.arxiv.org/abs/1911.05047
  32. [32]F. Natterer. The mathematics of computerized tomography. SIAM, 2001.
  33. [33]F. Natterer and F. Wübbeling. Mathematical Methods in Image Reconstruction, volume 5 of Monographs on Mathematical Modeling and Computation. SIAM, Philadelphia, PA, 2001.
  34. [34]Juan Peypouquet, Siqi Qu, and Mathias Staudigl. Asymptotic behavior of penalty dynamics for constrained variational inequalities. Applied Mathematics & Optimization, 93(2):36, 2026.arxiv.org/abs/2503.03902
  35. [35]I.W. Selesnick, R.G. Baraniuk, and N.C. Kingsbury. The dual-tree complex wavelet transform. IEEE Signal Processing Magazine, 22(6):123–151, 2005.DOI
  36. [36]Eduardo Soares, Plamen Angelov, Sarah Biaso, Michele Higa Froes, and Daniel Kanda Abe. SARS-CoV-2 CT-scan dataset: A large dataset of real patients ct scans for SARS-CoV-2 identification. medRxiv, 2020.DOI
  37. [37]Mikhail Solodov. An explicit descent method for bilevel convex optimization. Journal of Convex Analysis, 14(2):227, 2007.DOI
  38. [38]Mathias Staudigl and Siqi Qu. Tikhonov regularized exterior penalty dynamics for constrained variational inequalities. IEEE Control Systems Letters, 8:622–627, 2024.arxiv.org/abs/2403.13460
  39. [39]G. W. Stewart. On the perturbation of pseudo-inverses, projections and linear least squares problems. SIAM Review, 19(4):634–662, 1977.DOI
  40. [40]Thomas Strohmer and Robert W Heath Jr. Grassmannian frames with applications to coding and communication. Applied and computational harmonic analysis, 14(3):257–275, 2003.arxiv.org/abs/math/0301135
  41. [41]Joel A Tropp, Inderjit S Dhillon, Robert W Heath, and Thomas Strohmer. Designing structured tight frames via an alternating projection method. IEEE Transactions on information theory, 51(1):188–209, 2005.DOI
  42. [42]Shayne F. D. Waldron. An Introduction to Finite Tight Frames. Applied and Numerical Harmonic Analysis. Birkhäuser, New York, NY, 1 edition, 2018. 25 b/w illustrations, 12 colour illustrations.DOI
  43. [43]Nachuan Xiao, Xin Liu, and Ya-xiang Yuan. Exact penalty function for ℓ₂,₁ norm minimization over the stiefel manifold. SIAM Journal on Optimization, 31(4):3097–3126, 2021.DOI
  44. [44]Wei Hong Yang, Lei-Hong Zhang, and Ruyi Song. Optimality conditions for the nonlinear programming problems on riemannian manifolds. Pacific Journal of Optimization, 10(2):415–434, 2014.
  45. [45]Quan Zhang, Hang Wang, Wei Chen, and Guangtan Huang. A local radon transform for seismic random noise attenuation. Journal of Applied Geophysics, 186:104264, 2021.DOI

Paper details

vlnc:W_8ubxhdhn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents