Fact-checked by Grok 2 weeks ago

Multidimensional system

A multidimensional system is a mathematical framework in that models dynamic processes or signals as functions of multiple independent variables, such as time and , in contrast to one-dimensional systems governed by a single variable like time alone. These systems are prevalent in disciplines including , where they describe data structures like two-dimensional images or three-dimensional videos, and , where they handle spatially distributed or repetitive processes. Key aspects of multidimensional systems include their using partial or equations, with linear shift-invariant systems characterized by multidimensional s and transforms for and filtering. For instance, a two-dimensional signal x[n_1, n_2] can be processed via y[n_1, n_2] = \sum_{m_1} \sum_{m_2} x[m_1, m_2] h[n_1 - m_1, n_2 - m_2], enabling efficient operations like separable filtering that reduce from O(N^4) to O(2N^3). In applications, is more complex due to the multi-variable nature, often requiring tools like Gröbner bases or behavioral approaches to ensure proper controller design. The importance of multidimensional systems stems from their role in modern technologies, such as image reconstruction, array , and networked control systems, where handling high-dimensional data is essential for applications in healthcare, , and processing. Challenges include ensuring in higher dimensions and leveraging architectures like VLSI for efficient implementation. Ongoing focuses on innovative techniques, such as of lossless matrices, to advance and system performance.

Fundamentals

Definition and Scope

A , or m-D system with m > 1, refers to a whose inputs and outputs are functions of m variables, extending beyond the single-variable case of traditional one-dimensional systems. These variables typically represent spatial coordinates, time, or combinations thereof; for example, a two-dimensional () discrete signal for an might be denoted as x(i, j), where i and j are integer indices corresponding to row and column positions. In a three-dimensional () case, such as video data, the signal could be x(i, j, k), incorporating two spatial dimensions and a temporal one. The scope of multidimensional systems includes both continuous-time and discrete-time formulations, with signals defined over domains like \mathbb{R}^m for continuous cases or \mathbb{Z}^m for discrete lattices. The theory predominantly emphasizes linear shift-invariant (LSI) systems in the domain or linear time-invariant (LTI) systems in the continuous domain, where the output is obtained via multidimensional with an . Nonlinear multidimensional systems exist as extensions but receive less theoretical attention due to increased complexity. Distinctive characteristics of m-D systems arise from their higher-dimensional structure, including commuting shift operators across dimensions that nonetheless introduce partial orders for causality—unlike the total order in one dimension—complicating recursive implementations. Additionally, the lack of an analogue to the fundamental theorem of algebra prevents straightforward factorization of multivariate polynomials into linear factors, resulting in specialized challenges for stability testing (e.g., ensuring no zeros in the unit polydisk) and realization procedures.

Historical Development

The study of multidimensional systems began in the 1960s as an extension of one-dimensional signal processing techniques to accommodate signals varying in multiple dimensions, particularly for image filtering applications. Early contributions focused on adapting recursive filtering methods to two-dimensional data, with Thomas S. Huang's work on digital picture processing marking a foundational step in handling spatial signals computationally. This period saw the transition from analog to digital approaches, driven by advances in computing that enabled practical implementation of 2D filters for tasks like enhancement and restoration. Key milestones in the 1970s included the formalization of the two-dimensional , which provided essential tools for analyzing multidimensional signals in the z-domain, as explored in works like those by Huang and collaborators. Concurrently, state-space models were developed to describe the dynamics of multidimensional systems more comprehensively. Roesser's 1975 discrete state-space model generalized one-dimensional frameworks to two dimensions for linear image processing, introducing a local state representation partitioned along spatial axes. Shortly thereafter, Fornasini and Marchesini proposed their 1976 state-space realization theory for two-dimensional filters, offering algebraic criteria for and that extended to higher dimensions. Influential publications in the early 1980s synthesized these developments, with the book Multidimensional Digital Signal Processing by Dan E. Dudgeon and Russell M. Mersereau serving as a comprehensive reference for theory, transforms, and . The 1990s shifted emphasis toward computational efficiency, leveraging emerging hardware to implement multidimensional algorithms in applications like . In the post-2000 era, multidimensional systems have integrated with array processing techniques and sparse representations, enhancing efficiency in handling high-dimensional data such as sensor arrays and volumetric images. For instance, shearlet systems have emerged as powerful tools for sparse approximations in multidimensional domains, improving compression and analysis in signal processing tasks. More recently, from the 2010s to 2025, advances have included the incorporation of deep learning techniques for multidimensional signal analysis, such as convolutional neural networks for enhanced image and video processing, and topological signal processing for modeling complex data structures.

Applications

In Signal and Image Processing

In signal and image processing, multidimensional systems provide the theoretical foundation for handling signals indexed by multiple independent variables, such as spatial coordinates in images or spatiotemporal dimensions in videos, enabling efficient filtering and analysis of complex data structures. These systems extend one-dimensional (1D) signal processing techniques to higher dimensions, where operations like convolution and filtering must account for interactions across all indices to preserve signal integrity and avoid artifacts. For instance, two-dimensional (2D) finite impulse response (FIR) and infinite impulse response (IIR) filters are widely applied for image enhancement, edge detection, and restoration tasks, with FIR filters offering linear phase characteristics ideal for smoothing noise while preserving edges, and IIR filters providing computational efficiency through recursion for tasks like deblurring degraded images. A key technique in these applications is multidimensional convolution, which computes the output signal as a weighted sum over multiple dimensions, allowing filters to capture local patterns such as textures or gradients in images more effectively than separable 1D operations. In recursive (IIR) filters, quarter-plane is enforced to ensure computational feasibility and , restricting the filter's to the first of the index plane (non-negative shifts in both directions) and thereby preventing dependence on future samples that could introduce non-causal delays in . Extending to three dimensions, multidimensional systems are applied in for tasks such as and artifact reduction, exploiting correlations across spatial and temporal dimensions to achieve efficient . Post-2000 advancements have integrated multidimensional systems into frameworks, particularly convolutional neural networks (CNNs) for , where convolutional layers perform multidimensional convolutions to extract hierarchical features from images, enabling applications like with reduced parameters compared to fully connected networks. Similarly, multidimensional transforms facilitate multiresolution by decomposing signals into subbands at varying scales and orientations, supporting efficient and denoising in images while capturing both low- and high-frequency components. In practical examples, image processing employs 2D multidimensional filters for , enhancing multispectral data to delineate landforms and vegetation patterns for . In biomedical , 2D and multidimensional models aid MRI by iteratively applying filters to under-sampled data, improving resolution and reducing acquisition time without compromising diagnostic accuracy.

In Control and Other Engineering Fields

In , multidimensional systems are essential for modeling spatially distributed processes where dynamics evolve across multiple spatial dimensions, enabling the design of robust controllers for complex applications. For instance, two-dimensional diffusion systems, governed by partial equations (PDEs), are approximated using methods to derive finite-order models suitable for process control, allowing for reduced-order representations that maintain accuracy in temperature regulation tasks. Similarly, boundary control strategies for multidimensional equations in parallelepiped domains use Fourier methods and integral equations to achieve prescribed average temperatures, demonstrating the feasibility of admissible controls via Laplace transforms for industrial heating systems. In , multidimensional system models facilitate the control of antennas, where two-dimensional phased arrays employ mode diversity in few-mode fibers to perform with true time delays, supporting precise scanning without beam squint in applications like radars. Extending to other fields, biomedical imaging leverages multidimensional models in to characterize three-dimensional particle-discrete datasets, enabling of size, shape, and density distributions in biological tissues through high-resolution micro-CT reconstructions. In communications, two-dimensional uniform planar integrate hybrid techniques to generate directive beams, optimizing user scheduling and power allocation in multi-antenna systems. Furthermore, PDE-based simulations model multidimensional , such as two-dimensional Navier-Stokes equations, using to predict with high fidelity across varying Reynolds numbers. Recent advancements in the 2020s highlight multidimensional in autonomous vehicles, where three-dimensional combines , camera, and data in bird's-eye-view representations to enhance , achieving mean average precision scores up to 70.2 on benchmarks like nuScenes for path planning and . In , integrated dynamics models using dual quaternions govern multi-joint arms, with sliding mode controllers ensuring convergence of position and attitude errors under disturbances, applicable to manipulation tasks. A key challenge in these applications is handling non-separability in multidimensional loops, where coprime factorizations and address stabilization and H∞ , overcoming complexities absent in one-dimensional systems.

Mathematical Models

Comparison to One-Dimensional Systems

One-dimensional (1D) systems exhibit sequential causality, where the output at any time depends only on current and past inputs due to the total ordering imposed by the time axis. In contrast, multidimensional (m-D) systems lack a global temporal ordering, resulting in partial causality defined relative to the coordinate system; for instance, a two-dimensional (2D) signal is causal if its support is confined to the quarter-plane where both indices are non-negative, allowing computation in a specific direction but without the unidirectional flow of 1D systems. This partial ordering complicates system behavior, as shifts in coordinates or rotations can alter causality properties, unlike the invariant sequential nature in 1D. Recursion in 1D systems supports straightforward forward or backward implementations for filters, enabling efficient computation with well-established stability tests like Jury's criterion. In m-D systems, however, recursion demands partitioning the support into recursive and non-recursive regions to mitigate instability, as the absence of total order prevents simple sequential evaluation; for 2D recursive filters, this often involves quarter-plane or fan-shaped frequency partitioning, increasing design complexity and requiring separate stabilization steps not needed in 1D. Seminal work highlighted these challenges by deriving simplified stability conditions for 2D recursive filters, showing that poles must lie outside the unit bidisk without the 1D analog of a single stability circle. The in 1D is a single-variable facilitating pole-zero analysis via the (FVT) and clear regions of convergence (ROC) as annuli. Multidimensional z-transforms extend this to multivariate forms, such as the 2D case X(z_1, z_2) = \sum_{n_1=-\infty}^{\infty} \sum_{n_2=-\infty}^{\infty} x(n_1, n_2) z_1^{-n_1} z_2^{-n_2}, where poles and zeros form surfaces in rather than isolated points, complicating analysis without a direct FVT equivalent and leading to intersecting loci that do not cancel as in 1D. Computationally, the 1D (FFT) achieves O(N \log N) complexity for an N-point signal, enabling efficient frequency-domain processing. For m-D signals, the multidimensional FFT separates into 1D transforms along each , yielding O(N^d d \log N) complexity for a d-dimensional of size N per , which scales exponentially with dimensionality and often requires separable approximations to manage the increased load, as seen in applications where 2D FFTs process twice as costly per compared to 1D sequences.

Linear State-Space Models

Linear state-space models provide a time-domain framework for representing linear time-invariant (LTI) multidimensional systems, particularly useful for cases where signals are indexed by two discrete variables (i, j). In these models, the state is often described using local state vectors that account for shifts in different directions, such as horizontal and vertical for systems. This approach extends the classical 1D state-space representation by partitioning the state to reflect the multi-directional dependencies inherent in multidimensional signals. A foundational representation is the Roesser model, introduced for discrete 2D LTI systems in the context of image processing. It employs two local state vectors: x^h(i,j) \in \mathbb{R}^{n_1} for the horizontal direction and x^v(i,j) \in \mathbb{R}^{n_2} for the vertical direction, capturing the propagation along each axis. The model equations are given by: \begin{bmatrix} x^h(i+1,j) \\ x^v(i,j+1) \end{bmatrix} = \begin{bmatrix} A_{11} & A_{12} \\ A_{21} & A_{22} \end{bmatrix} \begin{bmatrix} x^h(i,j) \\ x^v(i,j) \end{bmatrix} + \begin{bmatrix} B_1 \\ B_2 \end{bmatrix} u(i,j) y(i,j) = \begin{bmatrix} C_1 & C_2 \end{bmatrix} \begin{bmatrix} x^h(i,j) \\ x^v(i,j) \end{bmatrix} + D u(i,j) where u(i,j) is the input, y(i,j) is the output, and the matrices have appropriate dimensions. Boundary conditions are specified as x^h(0,j) = f_1(j) and x^v(i,0) = f_2(i) for i,j \geq 0, enabling the handling of initial states on the axes. An alternative formulation is the Fornasini-Marchesini model, which uses a single x(i,j) \in \mathbb{R}^n and combines shifts from both directions in a coupled manner. The second-order version, widely adopted for its structural properties, is defined by: x(i+1,j+1) = A_1 x(i,j+1) + A_2 x(i+1,j) + B u(i,j) y(i,j) = C x(i,j) + D u(i,j) with boundary conditions x(i,0) = f_1(i) and x(0,j) = f_2(j). This model realizes a broader class of 2D filters and allows algebraic analysis of and . Extensions to n-dimensional (nD) systems generalize these models using tensor notations, where the state is partitioned into n components corresponding to each index direction. For the Roesser-type nD model, the update involves an n-partitioned and a block-structured , facilitating analysis in higher dimensions like 3D . The Fornasini-Marchesini nD variant similarly couples updates across multiple indices, though realizations grow complex with dimensionality. These generalizations maintain the core state-space structure but are primarily conceptual, with practical use limited by computational demands. State-space models like Roesser and Fornasini-Marchesini offer advantages in handling boundary and initial conditions explicitly, which is crucial for multidimensional systems where signals may have non-zero starts along multiple boundaries. They also enable computability through iterative operations, supporting simulations and design in applications such as filtering and stability analysis. These representations relate to frequency-domain tools via z-transforms but emphasize time-domain dynamics.

Multidimensional Transfer Functions

In multidimensional s, the provides a frequency-domain representation of the system's input-output relationship, generalizing the one-dimensional to multiple variables. For a two-dimensional () linear shift-invariant system, the output Y(z_1, z_2) relates to the input X(z_1, z_2) via the T(z_1, z_2) as Y(z_1, z_2) = T(z_1, z_2) X(z_1, z_2), where the multivariate is applied. For causal quarter-plane filters, common in 2D , the takes a rational form: T(z_1, z_2) = \frac{\sum_{p=0}^{M} \sum_{q=0}^{N} b_{p,q} z_1^{-p} z_2^{-q}}{1 + \sum_{p=0}^{M} \sum_{q=0}^{N} a_{p,q} z_1^{-p} z_2^{-q}}, where the numerator represents the zero-order autoregressive moving average (ARMA) part and the denominator the terms, with coefficients a_{p,q} and b_{p,q} defining the filter order. This structure ensures by restricting support to the quarter-plane where indices are non-negative. The can also be derived from state-space realizations, such as the Roesser model, which partitions the state into and vertical components. For this model, the transfer function is given by T(z_1, z_2) = C \begin{bmatrix} z_1 I_{n_1} - A_{11} & -A_{12} \\ -A_{21} & z_2 I_{n_2} - A_{22} \end{bmatrix}^{-1} \begin{bmatrix} B_1 \\ B_2 \end{bmatrix}, assuming no direct feedthrough term, where A_{ij}, B_i, and C are system matrices, and n_1, n_2 are state dimensions. This expression links time-domain state evolution to the frequency-domain behavior. In the n-dimensional (nD) case, the generalizes to a ratio of multivariate polynomials in z_1, \dots, z_n, often expressed as T(z_1, \dots, z_n) = \frac{N(z_1, \dots, z_n)}{d(z_1, \dots, z_n)}, where N and d are polynomials derived from state-space models like Fornasini-Marchesini. Computationally, separability assumptions—where polynomials factor into products of lower-dimensional terms—are frequently imposed to reduce complexity, enabling efficient realization via one-dimensional techniques. Key properties of multidimensional transfer functions include the absence of a simple zero-location test for , unlike the one-dimensional Jury test, necessitating analogs of the that involve solving linear matrix inequalities or frequency-domain conditions on the unit polydisk. These properties arise from the coupled nature of multiple variables, complicating pole-zero analysis.

Realization and Implementation

Structures for 2D Transfer Functions

The realization of two-dimensional (2D) transfer functions involves converting the rational polynomial expression into a state-space model, followed by transformation into practical hardware or software structures such as direct-form or cascade realizations. A common approach starts with the Roesser state-space model, which partitions the state vector into horizontal and vertical components to model recursive dependencies in both dimensions. From this state-space form, direct realizations can be derived by mapping the system matrices to delay elements and adders, while cascade structures decompose the overall transfer function into interconnected first-order sections for modularity. Partial fraction expansion, a standard technique in one dimension for simplifying realizations, faces significant challenges in 2D due to the lack of unique factorization in the ring of polynomials in two variables, often requiring approximations or iterative methods to achieve decomposable forms. Common structures for implementing transfer functions leverage the Roesser model for realization, incorporating delay elements that operate independently in the horizontal and vertical directions to handle quarter-plane . This model facilitates efficient mapping to digital hardware by representing shifts along each axis with unit delays, enabling of state updates. To mitigate , especially for higher-order systems, separable approximations are employed, where the transfer function is approximated as a product of one-dimensional factors in the horizontal and vertical variables, reducing the number of multipliers and delays required. These approximations preserve key characteristics while simplifying synthesis for applications. For (IIR) structures derived from all-pole transfer functions, loops are configured separately for horizontal and vertical recursions, using the Roesser model's partitioned state transitions to enforce and without cross-dimensional in the paths. In general cases with both poles and zeros, the structure extends to include paths that incorporate numerator terms, often realized through combinations of all-pole sections to non-minimum-phase behaviors. These IIR configurations are particularly suited for applications requiring selectivity, as the directional minimizes round-off noise compared to fully coupled alternatives. Computational realizations of 2D s frequently utilize the via the two-dimensional (2D DFT), where the input signal is transformed, multiplied pointwise by the filter's (evaluated from the ), and then inverse-transformed to yield the output. This approach is efficient for block-based processing, as fast algorithms like the 2D FFT reduce complexity from O(N^4) to O(N^2 log N) for an N x N . For real-time hardware implementation, field-programmable gate arrays (FPGAs) are widely adopted due to their reconfigurability and parallelism, allowing pipelined execution of delay and multiplication operations across multiple processing elements to handle high-throughput 2D filtering tasks such as enhancement.

Finite and Infinite Impulse Response Examples

Finite impulse response () filters in multidimensional systems, particularly two-dimensional () cases, are characterized by transfer functions that are polynomials in the inverse variables z_1^{-1} and z_2^{-1}, representing all-zero structures with no poles except at the origin. A general 2D transfer function takes the form T(z_1, z_2) = \sum_{p=0}^{M} \sum_{q=0}^{N} b_{p,q} z_1^{-p} z_2^{-q}, where b_{p,q} are the coefficients corresponding to the impulse response h(i,j), which has finite support limited to $0 \leq i \leq M and $0 \leq j \leq N for quarter-plane . This non-recursive nature allows for a straightforward state-space realization using the Roesser model, where the system A = 0, the input B incorporates components for delay lines in horizontal and vertical directions, and the output is a weighted sum of past inputs stored in the . Such realizations avoid , ensuring inherent . A numerical example of a simple 2D low-pass filter uses a 2x2 to approximate averaging over neighboring samples. The is T(z_1, z_2) = \frac{1}{4} (1 + z_1^{-1} + z_2^{-1} + z_1^{-1} z_2^{-1}), with coefficients b_{0,0} = 0.25, b_{1,0} = 0.25, b_{0,1} = 0.25, b_{1,1} = 0.25. The h(i,j) is nonzero only for (i,j) = (0,0), (1,0), (0,1), (1,1), producing a effect in image processing applications. In state-space form via the Roesser model, the and vertical states capture one-step delays, with A = \begin{bmatrix} 0 & 0 \\ 0 & 0 \end{bmatrix}, B = \begin{bmatrix} 1 \\ 1 \end{bmatrix} (adjusted for input ), and output matrix C selecting the weighted contributions, resulting in a minimal order of 2. In contrast, (IIR) filters in 2D systems feature all-pole transfer functions, enabling recursive computation but introducing potential instability. A all-pole example is T(z_1, z_2) = \frac{1}{1 + a_{1,0} z_1^{-1} + a_{0,1} z_2^{-1}}, where the coefficients a_{1,0} and a_{0,1} attenuation in the and vertical directions, respectively, with the h(i,j) extending infinitely in both indices. This can be realized using the Roesser model, partitioning the into (x^h) and vertical (x^v) components to handle coupled delays, incorporating terms that reflect the recursive dependencies. For instance, the model equations involve matrices where off-diagonal elements capture the cross-coupling between directions, such as A_{12} linking vertical state to horizontal update and A_{21} for the reverse. FIR realizations are stable by design due to the absence of loops and finite-duration response, though they require higher orders to achieve sharp selectivity, increasing computational demands. IIR structures, while more efficient with lower orders for equivalent , risk if the poles lie outside the unit polydisk, necessitating careful coefficient selection to satisfy criteria.

Analysis and Challenges

Stability Criteria

In multidimensional systems, bounded-input bounded-output (BIBO) stability requires that every bounded input produces a bounded output, which for two-dimensional (2D) discrete systems is equivalent to the impulse response h(i,j) being absolutely summable over the quarter-plane, i.e., \sum_{i=0}^\infty \sum_{j=0}^\infty |h(i,j)| < \infty. Unlike one-dimensional systems, where the Jury test provides a straightforward algebraic check, no simple analogous test exists for 2D BIBO stability, necessitating more complex frequency-domain or state-space approaches. Frequency-domain criteria for stability often involve ensuring that the denominator polynomial of the transfer function has no zeros inside or on the unit polydisk. For 2D polynomials, extensions of the use functional Schur coefficients derived from slice functions along the unit torus, where stability holds if all such coefficients satisfy |\gamma_k(\omega)| \leq 1 for k = 1 to the degree and all \omega on the unit circle. In state-space models, the provides a condition for BIBO stability by verifying that the H_\infty-norm of the transfer function satisfies \|T\|_\infty < 1, adapted to 2D or via lossless bounded-real conditions involving positive definite solutions to Lyapunov-like inequalities. For higher-dimensional (nD) systems, stability tests build recursively on lower-dimensional checks, addressing the exponential increase in complexity. For instance, 3D stability can be assessed by treating the system as a family of 2D systems parameterized by the third variable and verifying 2D stability conditions over the unit circle in that variable. The following table summarizes key nD stability tests:
DimensionTest NameDescriptionReference
2DHuang TestChecks no zeros in $z_1
2DBistritz Tabular FormConstructs a Jury-like table from polynomial coefficients to test stability without boundary zeros.
3DRecursive 2D SlicingFixes one variable on the unit circle and applies 2D tests recursively to resulting bivariate polynomials.
nDAnderson-Jury ExtensionGeneralizes to nD by ensuring no zeros in the unit polydisk via iterative 1D stability on hyperplanes.
Challenges in nD stability include the presence of non-minimum phase zeros, which, unlike in 1D systems, can destabilize the system even if poles are inside the unit polydisk due to the coupled variable dependencies. Post-2000 developments have introduced Lyapunov-based methods for nonlinear multidimensional systems, extending classical 1D Lyapunov theory to Roesser-form models. For continuous-time 2D nonlinear systems, asymptotic stability of the equilibrium requires a Lyapunov function V(x_h, x_v) = V_h(x_h) + V_v(x_v) such that its partial derivatives along trajectories satisfy \frac{\partial V_h}{\partial t_1} \leq -\gamma_h(\|x_h\|) and similarly \frac{\partial V_v}{\partial t_2} \leq -\gamma_v(\|x_v\|) for the vertical direction, where \gamma_h, \gamma_v are class K functions; this generalizes to nD via summed dimension-wise functions. Numerical tools, such as sum-of-squares optimization in MATLAB via SOSTOOLS, facilitate constructing these Lyapunov functions for polynomial nonlinearities by solving semidefinite programs. More recent advances as of 2024 include unified linear matrix inequality (LMI) conditions for 2D stability analysis and enhanced criteria for nD discrete state-space systems.

Factorization and Computational Issues

In multidimensional systems, the factorization of polynomials presents fundamental challenges that diverge sharply from one-dimensional cases. Unlike univariate polynomials, where the Fundamental Theorem of Algebra guarantees factorization into linear factors over the complex numbers, multivariate polynomials in m dimensions (m > 1) lack such an analog, as they do not necessarily factor into degree-one factors even over algebraically closed fields like the complexes. This absence stems from the geometry of zero sets in higher dimensions, which form unbounded hypersurfaces rather than isolated points, preventing the isolation of individual zeros as in 1D. For instance, in 2D systems, a polynomial like z_1 z_2 - 1 is irreducible over \mathbb{C}, with its zero set comprising a continuous hyperbola, illustrating how factors cannot be separated into isolatable roots. Consequently, most m-D polynomials (m > 1) are irreducible, with reducible ones forming a set of measure zero in the coefficient space, complicating tasks like greatest common divisor computation and minimal realizations. These factorization hurdles extend to computational issues in implementing multidimensional systems, particularly in and applications. Direct computation of m-dimensional convolutions, essential for filtering and system simulation, exhibits exponential complexity O(N^{2m}), where N is the support size per dimension, rendering exact evaluations infeasible for large-scale data even in or higher. To mitigate this, approximations leverage separability—decomposing kernels into products of 1D filters, reducing complexity to O(m N^{m+1})—or tensor decompositions like CANDECOMP/PARAFAC () or models, which exploit low-rank structure for efficient storage and computation in contexts such as multidimensional imaging. Recent advances include GPU-accelerated multidimensional fast transforms (m-D FFTs) via libraries like cuFFT, enabling real-time processing of high-dimensional signals in applications from to seismic analysis, with significant speedups (often 2-5x) over CPU implementations for volumes up to $1024^3, as of benchmarks. Specific algorithms address in design tasks, such as 2D spectral for (IIR) filters, which decomposes positive spectra into minimum-phase factors to ensure and . Early methods, like those extending 1D Hilbert transforms to quarter-plane filters, yield analytic factors but often require infinite order for exactness in , unlike the finite 1D case. For practical IIR design, these are combined with optimization: post-2010 solvers employ (SDP) to approximate spectral factors under magnitude constraints, formulating nonnegativity via sum-of-squares polynomials and solving via interior-point methods with polynomial-time complexity in filter order. This SDP approach has enabled robust IIR filters for applications like image enhancement. However, scalability remains limited for m > 2 without hybrid GPU-SDP implementations, highlighting ongoing gaps in handling regimes of the .

References

  1. [1]
    Multidimensional Systems and Signal Processing
    ### Summary of Multidimensional Systems and Signal Processing
  2. [2]
    [PDF] Lecture 5: Multidimensional Signal Processing
    Summary. Lecture 5: Multidimensional Signal Processing. Mark Hasegawa-Johnson. ECE 417: Multimedia Signal Processing, Fall 2021. Page 2. Signals. Fourier.
  3. [3]
    Multidimensional Systems - an overview | ScienceDirect Topics
    Multidimensional systems refer to complex data sets characterized by a large number of quantitative features, where the interplay of these features can lead ...
  4. [4]
    [PDF] Multidimensional Signal Processing - DTIC
    Feb 4, 2023 · a. Page 6. The need for multidimensional signal processing in manifold areas of applications is now well recognized. The important role of ...
  5. [5]
    [PDF] Multidimensional Signal and Color Image Processing using Lattices
    Function of several independent variables (space, time, etc.) Scalar or vector valued signals. Continuous-domain and discrete-domain signals (where sampling.
  6. [6]
    Multidimensional Systems Theory and Applications - SpringerLink
    Free delivery 14-day returnsAuthors: N. K. Bose. N. K. Bose. University Endowed Fener Fellow, USA Department of Electrical Engneering, The Pennsylvania State University, USA.
  7. [7]
    [PDF] The definition of causality of multidimensional signals
    Jan 1, 2001 · Notice, that the causality of multidimensional signals is relative in respect to the shift of the coordinate system and in respect to a rotation ...
  8. [8]
    Digital Picture Processing: Volume 1 | Guide books
    Jan 9, 2014 · This book covers topics such as: Introduction, Mathematical Preliminaries, Visual Perception, Digitization, Compression, Enhancement, ...
  9. [9]
    [PDF] two-dimensional digital filtering and its - uncorrelated error analysis
    Jul 15, 1975 · Prior to 1965, the implementation of the two- dimensional ... dix I describes the properties of the two-dimensional Z-transform needed.
  10. [10]
  11. [11]
    Multidimensional digital signal processing : Dudgeon, Dan E
    Apr 6, 2022 · This book, 'Multidimensional digital signal processing', covers digital signal processing techniques, digital electronics, and multidimensional ...
  12. [12]
    [PDF] Efficient 2D Linear-Phase IIR Filter Design and Application in Image ...
    Abstract—We present an efficient and novel proce- dure to design two-dimensional (2D) linear-phase IIR filters with less hardware resource. A 2D linear- ...Missing: restoration | Show results with:restoration
  13. [13]
    Image Enhancement by Local Operators and Two-Dimensional ...
    ... FIR and hR 2-D digital filters. Two methods are described in the following as typical examples: they are also suitable for the enhancement applications. — ...Missing: seminal | Show results with:seminal
  14. [14]
    2D Convolution in Image Processing - Technical Articles
    Nov 30, 2018 · This article provides an insight on 2-D convolution and zero-padding with respect to digital image processing.
  15. [15]
    Design of stable, causal 2-D digital filters using real coefficient 2-D ...
    Presented here is a method for the design of 2-D causal quarter-plane recursive digital filters with real coefficients and arbitrary magnitude with/without.
  16. [16]
    A Survey of Convolutional Neural Networks: Analysis, Applications ...
    Jun 10, 2021 · Fifth, the applications of 1-D, 2-D, and multidimensional convolution are covered. Finally, some open issues and promising directions for CNN ...
  17. [17]
  18. [18]
    Terrain and Atmosphere Classification Framework on Satellite Data ...
    The CNN model was dedicated to processing satellite images in a framework for terrain analysis. For this purpose, we modeled a solution that maintains data ...
  19. [19]
    3D MRI Reconstruction Based on 2D Generative Adversarial ... - NIH
    In this paper, we propose the use of two-dimensional super-resolution technology for the super-resolution reconstruction of MRI images.
  20. [20]
    Reduced order models for a two-dimensional heat diffusion system
    Abstract. In this paper, a two-dimensional heat diffusion system modelled by a partial differential equation (PDE) is considered. Finite order approximations ...Missing: 2D multidimensional
  21. [21]
    Boundary control problem for a heat equation in a multidimensional ...
    This paper investigates the boundary control problem for the heat conduction equation in a multidimensional domain.Missing: 2D | Show results with:2D
  22. [22]
    Two-dimensional phased array antenna beamforming system based on mode diversity
    ### Summary: 2D Phased Array Beamforming and Multidimensional Systems in Satellite Communications
  23. [23]
    Three-Dimensional Particle-Discrete Datasets - MDPI
    This collection underscores the transformative potential of 3D particle-discrete datasets acquired through X-ray tomography in advancing particle technology.
  24. [24]
    Joint Hybrid Beamforming and User Scheduling Design - arXiv
    Besides, uniform planar array (UPA) is widely used in satellite communication systems due to its potential of generating two-dimensional (2D) directive beams.Missing: multidimensional | Show results with:multidimensional
  25. [25]
    Multi-resolution partial differential equations preserved learning ...
    Jan 13, 2024 · These physical phenomena are usually governed by partial differential equations (PDEs) and can be simulated by solving these PDEs numerically ...
  26. [26]
    A Review of Multi-Sensor Fusion in Autonomous Driving - PMC
    Oct 1, 2025 · Multi-modal sensor fusion has become a cornerstone of robust autonomous driving systems, enabling perception models to integrate ...Missing: 2020s | Show results with:2020s
  27. [27]
    Integrated dynamics and control of space multi-joint robotic arm
    This paper presents research on the integrated dynamics and control methods of a multi-joint robotic arm system installed on a spatial inertial platform.Missing: multidimensional | Show results with:multidimensional
  28. [28]
    [PDF] Feedback Control for Multidimensional Systems and Interpolation ...
    Feedback stabilization and optimal control problems for the case of classical linear sys- tems have been much studied over the past several decades.Missing: non- separability
  29. [29]
    Chapter 1. Multi-Dimensional Signals and Systems - Digital Video ...
    There are some fundamental differences between the theory of one-dimensional and multi-dimensional signal processing that arise from the related facts that ...
  30. [30]
    Analytical Design of Two-Dimensional Filters and Applications in ...
    Stability of the two-dimensional recursive filters is also an important issue and is much more complicated than for 1D filters. For 2D filters, in general, it ...Missing: challenges | Show results with:challenges
  31. [31]
    [PDF] Stability of Two-Dimensional Recursive Filters - DTIC
    We discuss some aspects of the stability problem in two- dimensional recursive filtering. In particular, we derive a simplified version of a ...Missing: 1965 | Show results with:1965
  32. [32]
    [PDF] Two-Dimensional Systems and Z-Transforms - Elsevier
    In summary, a 2-D Z-transform is specified by its functional form X(z1,z2) and its convergence region Rx. Similar to the 1-D case, the Z-transform is simply ...Missing: 1977 | Show results with:1977
  33. [33]
    [PDF] Lecture 12 DFT in Multiple Dimensions
    An intuitive way to understand the calculation of the DFT in multiple dimensions is to formulate it as a DFT along the first dimension. (keeping all other ...
  34. [34]
    [PDF] Control Systems Analysis for the Fornasini-Marchesini 2D Systems ...
    Four decades have elapsed since the Fornasini-Marchesini state-space model was published and the general aim of this paper is to provide a review of the ...
  35. [35]
    Doubly-indexed dynamical systems: State-space models and ...
    Mar 20, 1978 · Doubly-indexed dynamical systems provide a state space realization of two-dimensional filters which includes previous state models.
  36. [36]
    An approach to multidimensional Fornasini–Marchesini state-space ...
    This paper proposes a new approach that can generate a low-order Fornasini–Marchesini (F–M) model realization for a given n -D system by taking into account ...
  37. [37]
    [PDF] 2-D Finite Impulse Response (FIR) Filters • Difference equation y(m, n)
    • Quarter plane - then s<r means: (s2 < r2) and (s1 < r1) and s 6= r ... 2-D IIR Filter Transfer Functions. • Transfer function in Z-transform domain is.
  38. [38]
    [PDF] Nirmal K. Bose - Applied Multidimensional Systems Theory
    In 1983, Dr. Bose published a pioneering book entitled Applied Multidimensional. Systems Theory [2]. The first edition of his book emphasized new research ...
  39. [39]
    Multidimensional state-space models: A comparative overview
    This paper provides a comparative survey of the existing multidimensional state-space models. The 2-Dimensional state-space models are first presented and ...
  40. [40]
    [PDF] zeroing the transfer matrix of the roesser model of 2–d linear systems
    Feb 15, 2023 · It is shown that the controllability of the Roesser model is invariant under state feedbacks and the observability under output feedbacks.
  41. [41]
    (PDF) Transfer Function Computation for Multidimensional Systems
    Aug 7, 2025 · A new algorithm is presented for the determination of the coefficients of an n-dimensional (n-D) transfer function. The n-D state-space system ...
  42. [42]
    [PDF] Generalizations and new proof of the discrete-time positive real ...
    transfer function has poles on the stability boundary. ... Hill, “Discrete-time lossless bounded real lemma for multidimensional separable denominator digital ...
  43. [43]
    A discrete state-space model for linear image processing
    Feb 1, 1975 · The linear time-discrete state-space model is generalized from single-dimensional time to two-dimensional space.
  44. [44]
    Positive realization in canonical form of the 2D Roesser type model
    A procedure for calculation of the positive realization in canonical form of a transfer function is given. Published in: Decision and Control, 1997., ...
  45. [45]
    [PDF] 2-D Partial Fraction Expansions and Minimal Commutative ...
    Theorem 3: A 2-D transfer function w(Z1, Z2) ER(Z1, Z2) is realizable by a commutative 2-D system if and only if it admits a partial fraction expansion of the ...Missing: challenges | Show results with:challenges
  46. [46]
    Roesser model representation of 2D periodic behaviors: the (2,2)
    In this paper we consider 2D single-input/single-output behaviors ... 2D Roesser state-space representations (or realizations) for such behaviors.
  47. [47]
    Realization of 2-D denominator-separable digital filter transfer ...
    Feb 11, 1991 · In this paper the problem of realizing 2-D denominator-separable digital filter transfer functions is considered for processing of real ...
  48. [48]
    (PDF) Approximation of 2-D separable denominator recursive filters
    After introducing a two-dimensional (2-D) model for the class of quarter-plane-causal, recursive and separable in denominator (CRSD) filters, ...<|control11|><|separator|>
  49. [49]
    Positive real control of two-dimensional systems: Roesser models ...
    This paper considers the problem of positive real control for two-dimensional (2-D) discrete systems described by the Roesser model and also discrete linear ...
  50. [50]
    On the Minimal State-Space Realizations of All-Pole and All-Zero 2 ...
    A simple method is presented for the minimal state-space representation, described by the Roesser model, of a 2-D all-pole and all-zero transfer function.
  51. [51]
    Realization structure for a class of 2D recursive digital filters with ...
    An efficient method is presented for the design of a class of 2-D recursive digital filters with separable denominators and non-separable numerator transfer ...
  52. [52]
    [PDF] DFT Domain Image Filtering - Electrical and Computer Engineering
    How many 1D DFT's? – M rows: M N-pt DFT's. N l. N M t DFT'. – N columns: N M-pt DFT's. – M=N: 2N N-pt DFT's. – Each N-pt DFT requires N2 operations total ...Missing: transfer | Show results with:transfer
  53. [53]
    2D Frequency Domain Filtering and the 2D DFT
    To implement this transform, we would like an analog of the FFT, which will let us quickly compute the coefficients of the transform. In fact, we can do better.
  54. [54]
    2D Digital Filter Implementation on a FPGA - UWSpace
    Aug 31, 2011 · It is not efficient to implement 2D digital filters using a single processor for real-time applications due to the large amount of data. A ...
  55. [55]
    Multiprocessor FPGA implementation of a 2D digital filter - IEEE Xplore
    ABSTRACT. High performance implementation of 2D digital filters are highly desired in many applications for real-time processing.
  56. [56]
    Two-Dimensional Digital Signal Processing I - SpringerLink
    Book Title · Two-Dimensional Digital Signal Processing I ; Book Subtitle · Linear Filters ; Editors · Thomas S. Huang ; Series Title · Topics in Applied Physics ; DOI ...
  57. [57]
  58. [58]
    [PDF] Stability Analysis for Two-Dimensional Systems
    Bounded-Output (BIBO) stability criteria of 2-D systems are given. Several checkable sufficient conditions are also described. For the special cases when n ...
  59. [59]
    [PDF] BIBO stability test based on functional Schur coefficients - HAL
    Oct 24, 2007 · A 2-D criterion for testing stability was given by Huang and further developed for the multidimensional case by Anderson and Jury and by ...
  60. [60]
  61. [61]
    [PDF] Stability testing of two-dimensional discrete linear system ... - CORE
    Abstract— A new test for determining whether a bivariate polynomial does not vanish in the closed exterior of the unit bicircle (is stable) is developed.
  62. [62]
    Lyapunov stability for continuous-time multidimensional nonlinear ...
    Aug 7, 2025 · This paper deals with the stability of continuous-time multidimensional nonlinear systems in the Roesser form. The concepts from 1D Lyapunov ...
  63. [63]
    Memory-accelerated parallel method for multidimensional fast ...
    Jun 2, 2022 · We propose an efficient GPU-accelerated multidimensional FFT library to achieve better performance in this paper.
  64. [64]
    Research on high performance FFT algorithm on GPUs
    Jun 11, 2025 · In the paper, we have developed a GPU-accelerated efficient parallel multidimensional FFT algorithm on the CUDA, we aim to provide an ...
  65. [65]
    cuFFT - NVIDIA Developer
    cuFFT is a library that provides GPU-accelerated Fast Fourier Transform (FFT) implementations to build apps across disciplines.
  66. [66]
  67. [67]
    Efficient 2D Linear-Phase IIR Filter Design and Application in Image ...
    Nov 8, 2015 · A 2D linear-phase FIR filter prototype is first designed using semidefinite programming (SDP). The prototype filter is then decomposed into modular structures ...
  68. [68]
    [PDF] Scalable Multi-node Fast Fourier Transform on GPUs - arXiv
    Feb 28, 2022 · Abstract. In this paper, we present the details of our multi-node GPU-FFT library, as well its scaling on Selene HPC system.<|control11|><|separator|>