Oblique Projection Algorithms

• Oblique Projection Algorithms are linear projection schemes that project vectors onto target subspaces along non-orthogonal directions, enabling flexible decomposition and constraint enforcement.
• They accelerate convergence in iterative solvers and facilitate model order reduction by precisely matching system properties where orthogonal methods fall short.
• They find broad applications in time series prediction, blind source separation, and eigenvalue computations, offering improved performance over traditional orthogonal projections.

Oblique projection algorithms are a class of linear transformation and projection schemes fundamental to computational mathematics, numerical linear algebra, signal processing, model order reduction, machine learning, and dynamical systems. These algorithms generalize the notion of orthogonal projection by projecting vectors onto a target subspace along directions that are not necessarily orthogonal, enabling greater flexibility in subspace selection, decoupling, and constraint enforcement. Their utility stems from the ability to accommodate non-orthogonal decompositions, leading to improved approximation properties, faster convergence in iterative methods, and the ability to optimize task-specific objective functions where classical orthogonality constraints are suboptimal or limiting.

1. Definition and Formalism

Oblique projectors are idempotent linear operators that project vectors onto a chosen subspace along another (non-orthogonal) complementary subspace. For vector spaces over ℝ or ℂ, let $A\subset\mathbb{K}^n$ be the target subspace of dimension $r$, and let $B$ be a supplementary subspace such that $\mathbb{K}^n=A\oplus B$, but not necessarily $A\perp B$. Given full-rank matrices $V,W\in\mathbb{K}^{n\times r}$ whose columns span $A$ and whose left nullspace is $B$, the oblique projector is defined as

$P = V(W^TV)^{-1}W^T$

where $W^TV$ is invertible. $r$0 has $r$1, range $r$2, and kernel $r$3. Orthogonal projectors are recovered when $r$4 (i.e., $r$5; $r$6 with $r$7 orthonormal). In general, $r$8.

A fundamental property is that oblique projections separate components along non-orthogonal decompositions, allowing distinct treatment of signal, noise, or other physical modes that do not align with orthogonal complement spaces (Yin et al., 2014, Mo et al., 2024).

2. Key Algorithms and Methodologies

2.1 Iterative Linear Solvers: Oblique Kaczmarz Methods

The Kaczmarz method projects iterates onto solution hyperplanes orthogonally, which can stagnate if the system matrix has highly correlated rows. Oblique-projection Kaczmarz algorithms (Li et al., 2021, Wang et al., 2021) select, at each iteration, a search direction $r$9 so that the update projects onto a hyperplane along a non-orthogonal direction. Specifically, for two hyperplanes defined by rows $B$0 and $B$1 of $B$2, the search direction is $B$3, and the update is

$B$4

This step annihilates two residuals per iteration, dramatically accelerating convergence when the matrix rows are nearly collinear.

Both deterministic and randomized variants exist, with convergence rates that strictly improve over their purely orthogonal counterparts, especially in the high-correlation regime (Li et al., 2021, Wang et al., 2021).

2.2 Model Order Reduction (MOR)

Oblique projection algorithms underpin several state-of-the-art model reduction schemes where reduced order models (ROMs) are constructed to satisfy interpolation or approximation conditions that cannot be enforced with orthogonal projection alone (Zulfiqar et al., 2021, Zulfiqar et al., 2022, Errico et al., 22 Mar 2026). For instance, in frequency-weighted H₂-optimal MOR, the reduced-order state-space matrices are given by

$B$5

with $B$6 obtained by solving coupled Sylvester or Lyapunov equations to match first-order optimality conditions associated with frequency or time-localized weighted H₂-norms (Zulfiqar et al., 2021, Zulfiqar et al., 2022).

Oblique projections also naturally accommodate relative error minimization, stability constraints, and structure preservation beyond what orthogonal projection schemes can guarantee (Zulfiqar et al., 2022, Errico et al., 22 Mar 2026).

2.3 Prediction and Latent Dynamics Extraction in Time Series

Where the identification of a predictable latent subspace is desired, oblique projections associate the observed data with an optimally predictive latent VAR process, with the dynamic and static subspaces not required to be orthogonal (Mo et al., 2023, Mo et al., 2024). Given measurement data $B$7, an oblique projector of the form $B$8 (with $B$9) is iteratively optimized to minimize one-step prediction error covariance, providing dynamical components that are more predictable than those obtained by Principal Component Analysis (orthogonal projection).

Expectation-maximization (EM) schemes efficiently alternate between estimating latent variables and optimizing the oblique projector and underlying VAR coefficients (Mo et al., 2023, Mo et al., 2024).

2.4 Nonlinear and Data-driven Settings

In data-driven model reduction to spectral submanifolds, oblique projection algorithms efficiently approximate the correct base point on invariant manifolds, even for systems with significant non-normality, by learning projection directions aligned with the tangent space of the slow subspace and its complement (Bettini et al., 27 Mar 2025). This allows correct pairing between full and reduced system trajectories and dramatically improves model accuracy relative to orthogonal projections.

Similarly, operator learning in the Koopman framework benefits from oblique projections by admitting Petrov–Galerkin test spaces distinct from the feature space, thereby improving generalization and capturing dynamics outside the training region (Uchida et al., 2023).

2.5 Signal Processing and Blind Source Separation

Oblique projection schemes are used for blind source separation, such as in relative transfer function (RTF) estimation in multi-source beamforming. Here, the oblique projector blocks the subspace spanned by known sources, and the desired RTF is obtained as the principal eigenvector of the projected covariance matrix after noise whitening and augmentation with additional orthogonal vectors (Gode et al., 6 Aug 2025).

2.6 Eigenvalue Problems and Contour-based Solvers

The FEAST algorithm and its non-Hermitian generalization employ oblique projection within the Petrov–Galerkin framework for solving generalized eigenvalue problems. After constructing a suitable contour-integral trial subspace, the oblique projection is used to extract eigenpairs by enforcing orthogonality of the residual to a test subspace (often the image of the trial subspace under the system matrix), yielding accurate and parallelizable solvers (Yin et al., 2014).

3. Comparative Analysis and Implementation

Oblique projection algorithms frequently outperform their orthogonal counterparts in tasks where the underlying subspace structure is non-orthogonal or where specific approximation/interpolation constraints must be enforced. In iterative solvers for linear systems, oblique Kaczmarz-type schemes exhibit faster linear convergence and reduced iterations, with the cost of slightly more computation per iteration due to the construction of the projection direction (Li et al., 2021, Wang et al., 2021). In model order reduction, oblique projections enable exact interpolation of primal/dual system properties and can be tuned for optimal performance in frequency bands or time intervals of interest, at computational cost determined by sparse Sylvester equation solves and biorthonormalization (Zulfiqar et al., 2021, Zulfiqar et al., 2022).

Oblique projection methods in time series, such as PredVAR, achieve latent dynamic variables with higher predictability and more accurate reconstruction than one-shot or PCA-based methods, leveraging iterative EM schemes with eigenvalue problems on the predictive covariance (Mo et al., 2023, Mo et al., 2024). In operator learning, oblique projection allows learning flexible test functions to adapt the operator to broader data distributions, resulting in superior long-horizon prediction and control (Uchida et al., 2023).

The main implementation challenges are the conditioning of the involved bases (necessitating regularization), and, in some applications, the iterative determination of suitable dual/test bases to meet problem-specific optimality or orthogonality constraints (Zulfiqar et al., 2021, Li et al., 2021, Mo et al., 2024).

4. Application Domains

Oblique projection algorithms are fundamental in:

• Linear system solvers (fast row-action methods)
• High-dimensional model reduction in control, mechanics, and fluid dynamics
• Blind source separation and array signal processing
• Dimensionality reduction and feature extraction in time series (PredVAR, dynamic latent variable extraction)
• Data-driven discovery of invariant subspaces and operators (Koopman operator learning, spectral submanifold reduction)
• Non-intrusive reduced-order modeling for nonlinear stochastic PDEs (oblique CUR decompositions)
• Eigenvalue computations for generalized and non-Hermitian problems
• Machine learning ensemble tree methods via sparse oblique projections in splits (Tomita et al., 2015)

Their role is especially prominent when standard orthogonal projections either bias the solution, restrict attainable accuracy, or fail to enforce critical constraints (e.g., matching moments, interpolation, decoupling of dynamical and noise subspaces).

5. Notable Algorithms and Their Properties

Name/Domain	Core Oblique Projection	Notable Properties/Tasks
KO/Kaczmarz (Li et al., 2021)	Row–row direction	Annihilates two residuals; fast for correlated systems
TLRHMORA (Zulfiqar et al., 2022)	Gramian-driven bases	Time-limited, relative H₂ error minimization; no large Lyapunov equations
GasNiTROM (Errico et al., 22 Mar 2026)	Matrix manifold opt.	Stable latent dynamics, globally asymptotically stable, adjoint-based gradient
FEAST (non-Hermitian) (Yin et al., 2014)	Petrov–Galerkin	Accurate inner-contour eigenvalues; parallelizable, robust for singular pencils
PredVAR (Mo et al., 2024)	Predictability-ordered	EM-estimated optimal oblique loadings; maximally predictable latent variables
SSM reduction (Bettini et al., 27 Mar 2025)	Data-driven, foliation	Matches manifold base points for non-normal systems, improves backbone curves
RTF estimation (Gode et al., 6 Aug 2025)	Blocked subspace	Blocks previous sources, robust to noise via covariance whitening/subtraction

6. Open Challenges and Future Directions

Several research directions remain active:

• Ensuring global convergence in iterative oblique schemes, as stationarity is not always guaranteed (Zulfiqar et al., 2022, Zulfiqar et al., 2021).
• Extending oblique-projection invariance to block, nonlinear, and non-Euclidean geometric contexts.
• Development of efficient, scalable algorithms for very large-scale systems via randomized and sketch-based oblique projections (Donello et al., 2023).
• Application of oblique projection frameworks in reinforcement learning, especially unified views of temporal-difference and Bellman-residual methods (Scherrer, 2010).
• Automatic selection of dual/test subspaces in fully data-driven and learning-based contexts.

7. Comparative Performance and Practical Recommendations

When the underlying structure of the problem (e.g., non-orthogonality, asymmetric interpolation, correlated noise) suggests a non-orthogonal decomposition yields better performance, oblique projection algorithms provide superior flexibility and, in many cases, optimality. They should be implemented with appropriate attention to numerics, notably regularization of projectors, conditioning of test/trial bases, and robust eigen/singular value computations (Zulfiqar et al., 2022, Li et al., 2021, Mo et al., 2024). In time series modeling and model order reduction, oblique approaches (EM-based, Petrov–Galerkin) consistently yield latent representations with higher predictability and/or controlled error characteristics relative to variance-oriented (PCA) or one-shot techniques.

Oblique projection algorithms thus constitute a versatile and powerful methodological paradigm for a wide range of modern computational, learning, and signal processing tasks.