Deep Learning Assisted Subspace Methods

• Deep learning assisted subspace methods are hybrid approaches that use neural networks to restore or learn the geometric structures essential for reliable subspace estimation.
• They enhance classical techniques in array processing, clustering, and regression by constructing surrogate covariance matrices and latent self-expression maps while maintaining basis invariance.
• These methods overcome practical challenges like noise, coherence, and incomplete data by integrating learning with a model-based inference that preserves interpretable subspace geometry.

Deep learning assisted subspace methods are hybrid procedures in which a deep model is used to learn a representation, covariance surrogate, self-expression matrix, or subspace-valued map so that an explicitly subspace-based downstream method remains applicable. In the cited literature, this design appears in array processing for direction-of-arrival estimation, in deep subspace clustering, in principal-subspace and active-subspace discovery, and in parameter-to-subspace regression on Grassmann manifolds (Shmuel et al., 2023, Peng et al., 2017, Tripathy et al., 2019, Fanaskov et al., 27 Sep 2025). A recurrent theme is that the deep component does not simply replace the classical method with a black-box predictor; rather, it attempts to restore or learn the geometric structure that the subspace method requires.

1. Conceptual scope and taxonomy

Classical subspace methods assume that the data admit a decomposition or representation that is stable under a specific geometry. In direction-of-arrival estimation, subspace methods such as MUSIC and Root-MUSIC rely on a covariance matrix that can be divided into distinct signal and noise subspaces, and their performance degrades when assumptions such as narrowband signals, non-coherent sources, array calibration, enough snapshots, or a correct statistical model fail (Shmuel et al., 2023). In subspace clustering, sparse or low-rank self-expression assumes that samples from the same class lie in the same linear subspace; deep extensions were proposed precisely because many real datasets do not satisfy this assumption in the original space (Peng et al., 2017). In active subspaces, the classical formulation requires gradient information from the forward model, which is often impossible to obtain (Tripathy et al., 2019).

These works collectively define the practical scope of deep learning assistance: the network is introduced where the classical pipeline is fragile, ill-posed, or computationally expensive. The learned object varies by problem class, but the downstream target remains recognizably subspace-based.

Family	Learned object	Representative papers
DoA and sparse-array processing	Surrogate covariance or signal/noise subspaces	(Shmuel et al., 2023, Amiel et al., 2023, Chen et al., 2024)
Subspace clustering	Latent embedding, self-expression, or factorized affinity	(Peng et al., 2017, Zhang et al., 2018, Wu et al., 2023, Mrabah et al., 24 Dec 2025)
Grassmannian, principal, and active subspaces	Orthonormal projection, principal subspace, or subspace-valued regressor	(Lan et al., 2022, Tripathy et al., 2019, Fanaskov et al., 27 Sep 2025)

A central implication is that “subspace” is not used in a single narrow sense. In some papers it denotes signal and noise eigenspaces of a covariance operator; in others it denotes a union of linear subspaces used for clustering; in others it denotes a point on a Grassmann manifold or an active low-dimensional projection. What unifies the field is the attempt to preserve basis-invariant geometric structure while using deep learning to overcome the brittleness of purely model-based estimation.

2. Recurrent architectural pattern

A common architecture is a learned front end followed by a model-based subspace stage. SubspaceNet exemplifies this pattern in array processing: instead of predicting angles directly, it learns a covariance-like representation from empirical autocorrelations, reshapes the output into a complex matrix $K\in\mathbb{C}^{N\times N}$, and enforces Hermitian positive definiteness through

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$

so that the result remains suitable for eigendecomposition and Root-MUSIC (Shmuel et al., 2023). The same paper states explicitly that the network is trained through the differentiable Root-MUSIC block rather than against a ground-truth covariance matrix, because such a “decomposable covariance label” is generally unavailable.

A second pattern is latent self-expression. DSSC learns a nonlinear latent space $\mathbf{H}^{(M)}$ and then imposes the classical self-expressive relation

$$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$

with sparsity on $\mathbf{C}$ and a unit-sphere regularizer on the latent features (Peng et al., 2017). AMVDSN, DSASC, and MLRDSC preserve the same idea while altering the representation stage: attention for multi-view fusion, separate content and structure branches, or multi-level encoder features are all used to construct a latent space in which self-expression is more subspace-preserving (Lu et al., 2021, Wu et al., 2023, Kheirandishfard et al., 2020).

A third pattern is explicit subspace-valued learning. In sparse linear array localization, a DNN is trained to output a representation whose Gram matrix is eigendecomposed to obtain a basis for a $k$-dimensional subspace on a union of Grassmannians, and the loss is defined by principal angles and geodesic distance rather than by covariance matching (Chen et al., 2024). In deep active subspaces, the orthonormal projection matrix is parameterized as $W=h(Q)$, where $Q$ is unconstrained and $h(\cdot)$ applies Gram–Schmidt, so that ordinary backpropagation can optimize an orthonormal subspace projector (Tripathy et al., 2019). In subspace regression, the regressor itself outputs a subspace, and the loss is defined through projectors or stochastic surrogates that are invariant under basis changes (Fanaskov et al., 27 Sep 2025).

The significance of this pattern is methodological rather than cosmetic. These models try to learn quantities that remain meaningful under subspace equivalence, such as projectors, principal angles, or self-expression relations, instead of forcing the network to fit one arbitrary matrix representative.

3. Array processing and direction-of-arrival estimation

In direction-of-arrival estimation, deep learning assistance has been used primarily to repair the covariance geometry required by subspace methods. SubspaceNet begins from the standard ULA model

$X=A(\theta)S+V,$

and from the covariance decomposition

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$0

which underlies MUSIC and Root-MUSIC (Shmuel et al., 2023). The paper states that classical performance degrades when the standard assumptions fail, and proposes to learn an empirical autocorrelation-to-covariance map that restores a usable signal/noise split while preserving the interpretability and suitability of classical subspace estimators.

The sparse-array extension is given by Sparse-SubspaceNet, which is proposed for sparse miscallibrated arrays with coherent sources. Its abstract states that sparse arrays traditionally reconstruct the covariance of a virtual large ULA and then process it by subspace DoA estimators, but that these methods assume non-coherent signals and calibrated arrays. Sparse-SubspaceNet instead learns from data how to compute a surrogate virtual array covariance that is divisible into distinguishable subspaces, thereby learning to cope with coherent sources and miscalibrated sparse arrays while preserving interpretability (Amiel et al., 2023).

A more geometric reformulation appears in the sparse linear array literature. “Subspace Representation Learning for Sparse Linear Arrays to Localize More Sources than Sensors: A Deep Learning Methodology” argues that covariance reconstruction solves a harder problem than necessary because DoA estimation only needs the signal/noise subspace, not the whole covariance (Chen et al., 2024). The network is trained on $\hat{R}=KK^\mathsf{H}+\epsilon I_N,$1, uses losses derived from principal angles,

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$2

and employs the geodesic distance

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$3

The same paper proposes consistent rank sampling so that each minibatch contains the same source number $\hat{R}=KK^\mathsf{H}+\epsilon I_N,$4, making training over different Grassmann dimensions practical (Chen et al., 2024).

These DoA papers share a precise technical position: the deep network is not primarily an angle regressor. It is a learned intermediary that restores the algebraic conditions under which eigenspace separation, Root-MUSIC, or related estimators become reliable again. This suggests a stable division of labor between learning and model-based inference: the network estimates a subspace-compatible object, and the classical estimator still performs the final geometrically interpretable step.

4. Deep subspace clustering and self-expression

Deep subspace clustering is the largest concentration of deep-learning-assisted subspace methods in the cited corpus. DSSC is an early formulation that keeps Sparse Subspace Clustering’s self-expressive sparse-representation principle but moves it into a deep latent space (Peng et al., 2017). Its optimization combines latent reconstruction,

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$5

with the unit-sphere regularizer

$\hat{R}=KK^\mathsf{H}+\epsilon I_N,$6

and constructs the affinity matrix as $\hat{R}=KK^\mathsf{H}+\epsilon I_N,$7 before spectral clustering (Peng et al., 2017).

Subsequent work diversified the representation stage without abandoning self-expression. MLRDSC inserts multiple fully connected linear connection layers between encoder layers and their corresponding decoder layers so that self-expression is enforced at multiple encoder depths through a shared matrix $\hat{R}=KK^\mathsf{H}+\epsilon I_N,$8 and layer-specific matrices $\hat{R}=KK^\mathsf{H}+\epsilon I_N,$9 (Kheirandishfard et al., 2020). AMVDSN learns a joint latent representation that explicitly considers both consensus and view-specific information among multiple views, and then performs subspace clustering on the learned joint latent representation; it uses attention to derive dynamic weights for views and a Frobenius-regularized self-representation layer $\mathbf{H}^{(M)}$0 (Lu et al., 2021). DSASC uses a vision transformer to extract features, divides them into structure features and content features, learns separate self-expression matrices $\mathbf{H}^{(M)}$1 and $\mathbf{H}^{(M)}$2, and fuses them by $\mathbf{H}^{(M)}$3 for spectral clustering (Wu et al., 2023).

Another line focuses on direct subspace updates rather than affinity construction. Scalable Deep $\mathbf{H}^{(M)}$4-Subspace Clustering jointly learns a convolutional autoencoder and $\mathbf{H}^{(M)}$5 latent subspaces $\mathbf{H}^{(M)}$6, minimizing

$\mathbf{H}^{(M)}$7

and explicitly avoids the affinity-matrix bottleneck of self-expressive methods (Zhang et al., 2018). SDSNet later pursued the same scalability problem through landmark-based approximation and a factorized affinity $\mathbf{H}^{(M)}$8, with the paper claiming $\mathbf{H}^{(M)}$9 complexity when the landmark count is fixed and much smaller than the sample count (Mrabah et al., 24 Dec 2025).

Low-rank rather than sparse formulations also fit the deep-learning-assisted pattern. DeepLRR builds a multilayer collaborative low-rank coding network in which each layer decomposes the current representation into a deep principal feature part, a deep salient feature part, and a deep sparse error; the reconstructed output is then used as the input and low-rank dictionary for the next layer (Li et al., 2019). Energy-based self-expressiveness offers yet another modification: rather than storing self-expressive coefficients as a fully connected layer that requires the whole dataset in one batch, it learns an energy-based network to obtain the coefficients by mini-batch training and thereby allows deeper backbones such as ResNet (Li et al., 2021).

The field-wide consequence is that “deep subspace clustering” no longer refers to one fixed architecture. It names a family of models that differ in how they build the latent representation or the affinity, but preserve the belief that clustering is fundamentally recovered from subspace structure rather than from a direct class-probability head.

5. Grassmannian, principal, and active subspace learning

A more geometric branch of the literature treats the subspace itself as the learning target. The sparse linear array methodology already makes this explicit by training on a union of Grassmannians and proving that a ReLU network can approximate signal subspaces up to arbitrary average geodesic error (Chen et al., 2024). The same basis-invariant logic appears in “Deep Learning for Subspace Regression,” which formulates parameter-dependent reduced-order modeling as

$$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$0

and trains a neural network with subspace-aware losses such as

$$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$1

That paper further proposes predicting a larger-than-needed subspace, $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$2, and reports that for a 2D elliptic eigenproblem the test error for predicting the first 10 eigenvectors dropped from about $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$3 to about $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$4 when the model predicted a larger 40-dimensional subspace (Fanaskov et al., 27 Sep 2025).

Principal-subspace learning from sparse observations is addressed by Danskin-LISSA. The problem is to learn the $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$5-dimensional principal subspace of a matrix from sample entries rather than full columns, with the representation itself allowed to be a neural network (Lan et al., 2022). The paper defines a representation loss whose minimizer is the desired principal subspace, derives the gradient

$$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$6

and uses two independent LISSA estimates to construct a bias-controlled gradient estimator. The significance is that subspace learning becomes compatible with backpropagation even when only small random submatrices are observed (Lan et al., 2022).

Deep active subspaces occupy a related but distinct position. Classical active subspaces require gradients of the forward model, but the deep formulation replaces constrained optimization on the Stiefel manifold with the differentiable reparameterization $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$7, and learns the reduced surrogate

$$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$8

end to end (Tripathy et al., 2019). The paper reports that the learned subspace visually agrees well with the classical gradient-based active subspace on a stochastic elliptic PDE benchmark, even though the classical method performs better numerically (Tripathy et al., 2019).

Across these papers, basis invariance is not a secondary detail. It determines loss design, parameterization, and theoretical statements. A plausible implication is that deep learning becomes most compatible with subspace methodology when the optimization target is formulated in terms of projectors, principal angles, or inclusion of one subspace in another, rather than in terms of one arbitrarily chosen basis matrix.

6. Applications, limitations, and terminological boundaries

The application range is broad. In hyperspectral image clustering, SSC-friendly deep dictionary learning uses a deep nonlinear transformation based on deep dictionary learning and incorporates an SSC loss so that the transformed representation is separable into subspaces; the final affinity is $$\mathbf{H}^{(M)} \approx \mathbf{H}^{(M)}\mathbf{C},$$9 followed by normalized cuts (Goel et al., 2021). In uncertainty quantification, deep active subspaces provide gradient-free active-subspace recovery for high-dimensional surrogate modeling (Tripathy et al., 2019). In reduced-order modeling and numerical linear algebra, subspace regression is used for parametric eigenproblems, deflation techniques, relaxation methods, optimal control, and solution of parametric partial differential equations (Fanaskov et al., 27 Sep 2025). In robotics, deep neural network based subspace learning of manipulator workspace mapping learns specialized workspace subspaces from about $\mathbf{C}$0 samples, reports runtime reduction from $\mathbf{C}$1 s to $\mathbf{C}$2 s, and reports average F-measure $\mathbf{C}$3 with batch gradient descent and resilient backpropagation (Liao, 2018).

The limitations are equally recurrent. SubspaceNet is supervised, needs labeled training data, is trained for a given array/setup, and adds computational cost compared to pure model-based processing (Shmuel et al., 2023). The sparse linear array methodology still requires source-number conditioning and specialized batching across ranks (Chen et al., 2024). Deep subspace clustering methods often inherit the $\mathbf{C}$4 burden of full affinity construction and spectral decomposition, which is why scalable alternatives such as $\mathbf{C}$5SCN and SDSNet were proposed (Zhang et al., 2018, Mrabah et al., 24 Dec 2025). In hyperspectral clustering, the DDL-based formulation is nonconvex and shows performance degradation for larger windows, which the authors attribute to overfitting (Goel et al., 2021). In workspace mapping, performance degrades as workspace resolution increases or when the task is not aligned with the learned subspace (Liao, 2018).

A frequent misconception is that these methods are merely deep regressors with subspace terminology attached. The cited papers consistently contradict that reading. SubspaceNet preserves MUSIC/Root-MUSIC; DSASC still ends with spectral clustering on a fused self-expression matrix; subspace regression uses losses that vanish when the target subspace is contained in the predicted one; deep active subspaces explicitly enforce orthonormality of the active directions (Shmuel et al., 2023, Wu et al., 2023, Fanaskov et al., 27 Sep 2025, Tripathy et al., 2019). The deep component is therefore auxiliary to a subspace geometry that remains explicit.

The term also has an adjacent meaning in optimization, where the subspace is used to train the deep model rather than to assist a downstream inference procedure. Krylov Subspace Descent builds a Krylov subspace from the gradient and Hessian or Gauss–Newton products and then optimizes the network parameters inside that subspace (Vinyals et al., 2011). A two-stage subspace trust-region method for feed-forward neural nets similarly constructs a quadratic approximation in a low-dimensional subspace and performs first a positive-curvature trust-region step and then a gradient step (Dudar et al., 2018). These works are not subspace methods for clustering, DoA, or reduced-order modeling; they are subspace methods for deep-learning optimization. Their presence in the literature broadens, rather than narrows, the meaning of the phrase.

Taken together, the literature presents deep learning assisted subspace methods as a technically coherent but application-diverse research area. The deepest commonality is not the use of any particular architecture; it is the decision to keep the subspace object central, interpretable, and geometrically constrained while assigning to deep learning the task of making that object estimable under nonlinear, noisy, incomplete, or computationally prohibitive conditions.