Subspace Mapping (SuMa) Overview

• Subspace Mapping (SuMa) is a methodological family that projects high-dimensional objects onto lower-dimensional subspaces for optimal approximation and reduced computational complexity.
• It employs linear algebra techniques, such as pseudoinverses, to achieve efficient matrix approximations and subspace alignments in applications like domain adaptation, MRI reconstruction, and quantum computing.
• SuMa's versatility spans from teleoperation and deep learning to concept erasure in diffusion models, offering targeted solutions for alignment, regression, and robust feature neutralization.

Searching arXiv for recent and foundational papers on "Subspace Mapping" and closely related usages of the term. The literature suggests that Subspace Mapping (SuMa) is best understood not as a single standardized algorithm, but as a recurring research pattern in which a full-space object is replaced by, projected onto, or coordinated through a lower-dimensional or prescribed subspace. In linear algebra, this takes the form of approximating an observed matrix by a mapping from a prescribed subspace ${\mathcal T}$ onto a prescribed subspace ${\mathcal S}$, with optimal kernel $X = B^{\dag}AC^{\dag}$ and optimal approximation $BB^{\dag}AC^{\dag}C$ (Howlett et al., 2024). Closely related uses appear in unsupervised domain adaptation, quantitative MRI, teleoperation, diffusion-model concept erasure, quantum linear algebra, and multipartite entanglement, where the common motif is to encode, align, constrain, or neutralize information by operating on subspaces rather than on the ambient space directly (Fernando et al., 2014, Roeloffs et al., 2018, Meeker et al., 2019, Nguyen et al., 6 Sep 2025, Kerenidis et al., 2022, Marconi et al., 2 Apr 2025).

1. Scope and unifying structure

Across the cited literature, subspace mapping is organized around a small set of recurring ingredients: a source representation, a target representation, and a rule that either projects one onto the other or uses an intermediate subspace as a common coordinate system. In some cases the mapping is explicitly linear and closed-form; in others it is learned, regularized, or defined implicitly through optimization on the Grassmann manifold. This suggests that SuMa is a methodological family whose members differ by task, geometry, and optimization criterion rather than by a single canonical derivation.

Domain	Core subspace object	Representative formulation
Matrix approximation	Prescribed domain/range subspaces	$BB^{\dag}AC^{\dag}C$ (Howlett et al., 2024)
Domain adaptation	Source and target eigenspaces	$M^* = X_S'X_T$ (Fernando et al., 2014)
MRI reconstruction	Signal subspace of Bloch responses	Reconstruction and mapping performed entirely in subspace (Roeloffs et al., 2018)
Teleoperation	Shared 3D intermediary subspace	Spread, size, curl coordinates (Meeker et al., 2019)
Diffusion erasure	Target and reference concept subspaces	Target subspace mapped to reference subspace (Nguyen et al., 6 Sep 2025)
Quantum/entanglement	Subspace states or symmetric-state subspaces	Subspace states and multipartite-to-bipartite mapping (Kerenidis et al., 2022, Marconi et al., 2 Apr 2025)

A common misconception is that “SuMa” names one framework with stable notation across fields. The cited record suggests otherwise. In some papers the phrase denotes an exact projection problem on prescribed linear subspaces, in others a low-dimensional control intermediary, and in others a learned mechanism for concept neutralization or parameter-to-subspace regression. What remains stable is the preference for subspace-valued structure over unconstrained full-space representations.

2. Optimal approximation on prescribed subspaces

In the linear-algebraic formulation, an observed matrix $A \in {\mathbb C}^{m \times n}$ is approximated by a sum

$\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$

where the matrices $B_i \in {\mathbb C}^{m \times g_i}$ and $C_j \in {\mathbb C}^{h_j \times n}$ are known, and the unknown kernels ${\mathcal S}$0 are chosen by minimizing the Frobenius norm of the error. The sum can be represented as a bounded linear mapping ${\mathcal S}$1 with unknown kernel ${\mathcal S}$2 from a prescribed subspace ${\mathcal S}$3 onto a prescribed subspace ${\mathcal S}$4, defined respectively by the collective domains and ranges of the given matrices ${\mathcal S}$5 and ${\mathcal S}$6 (Howlett et al., 2024).

The central result is that the optimal kernel is

${\mathcal S}$7

and the optimal approximation is

${\mathcal S}$8

The latter is described as the projection of the observed mapping ${\mathcal S}$9 onto a mapping from $X = B^{\dag}AC^{\dag}$0 to $X = B^{\dag}AC^{\dag}$1 (Howlett et al., 2024). In the accompanying exposition, the same idea is phrased in basis form: if $X = B^{\dag}AC^{\dag}$2 and $X = B^{\dag}AC^{\dag}$3 encode the prescribed subspaces, then the best-fit subspace mapping uses the Moore–Penrose pseudoinverses and can be reduced to operations on much smaller matrices. The reduced form is given as

$X = B^{\dag}AC^{\dag}$4

together with the observation that one never needs to compute $X = B^{\dag}AC^{\dag}$5 or $X = B^{\dag}AC^{\dag}$6 explicitly at full scale when $X = B^{\dag}AC^{\dag}$7.

The computational significance is explicit. If $X = B^{\dag}AC^{\dag}$8 is large, then $X = B^{\dag}AC^{\dag}$9 and $BB^{\dag}AC^{\dag}C$0 may also be large, and direct calculation of $BB^{\dag}AC^{\dag}C$1 and $BB^{\dag}AC^{\dag}C$2 becomes unwieldy and inefficient. The proposed reduction method avoids this by reducing the solution process to finding the pseudo-inverses of a collection of much smaller matrices, which significantly reduces the computational burden (Howlett et al., 2024). This version of SuMa is therefore both an approximation principle and a matrix-reduction strategy.

3. Alignment, regression, and task-optimized embeddings

A closely related construction appears in Subspace Alignment for unsupervised domain adaptation. There, source and target domains are represented by low-dimensional subspaces $BB^{\dag}AC^{\dag}C$3, typically obtained by PCA, and one seeks a linear mapping $BB^{\dag}AC^{\dag}C$4 such that the mapped source subspace $BB^{\dag}AC^{\dag}C$5 is as close as possible to the target subspace $BB^{\dag}AC^{\dag}C$6 in Frobenius norm: $BB^{\dag}AC^{\dag}C$7 Because the columns are orthonormal, the solution has the closed form

$BB^{\dag}AC^{\dag}C$8

and the aligned source basis is

$BB^{\dag}AC^{\dag}C$9

The same paper also provides two approaches for choosing the subspace size: a theoretical stability bound and a maximum likelihood estimation procedure for high-dimensional data (Fernando et al., 2014). In this setting, subspace mapping is an alignment operator between latent eigenspaces rather than a projection of a single observed matrix.

The learning perspective is extended further in Subspace Support Vector Data Description, where data are mapped to a subspace optimized specifically for one-class classification. In the linear case, the projection is

$BB^{\dag}AC^{\dag}C$0

and the method jointly optimizes the mapping $BB^{\dag}AC^{\dag}C$1 and the hypersphere description of the target class. The optimization alternates between standard SVDD in the projected space and gradient-based updates of $BB^{\dag}AC^{\dag}C$2, with orthogonality enforced by $BB^{\dag}AC^{\dag}C$3 and regularization terms of the form

$BB^{\dag}AC^{\dag}C$4

(Sohrab et al., 2018). Here, the mapped subspace is not prescribed a priori; it is learned so as to make the target class more compactly enclosable.

In Deep Learning for Subspace Regression, the object being learned is itself a subspace-valued function. A parameter $BB^{\dag}AC^{\dag}C$5 is mapped to a linear subspace of $BB^{\dag}AC^{\dag}C$6, represented on the Grassmann manifold. Given training pairs $BB^{\dag}AC^{\dag}C$7, the model seeks

$BB^{\dag}AC^{\dag}C$8

with losses that depend only on subspaces rather than on their specific bases. Two losses are proposed: a projector-based loss

$BB^{\dag}AC^{\dag}C$9

and a stochastic least-squares-based loss $M^* = X_S'X_T$0, whose expectation equals $M^* = X_S'X_T$1 (Fanaskov et al., 27 Sep 2025). A distinctive device is subspace embedding, where the network predicts a larger subspace $M^* = X_S'X_T$2 instead of the minimal target subspace. The paper states that this decreases the complexity of the mapping for elliptic eigenproblems with constant coefficients and makes the mapping smoother for general smooth functions on the Grassmann manifold (Fanaskov et al., 27 Sep 2025). This shifts SuMa from closed-form linear algebra toward learned subspace-valued regression.

4. Subspace-constrained reconstruction and intermediary control

In quantitative MRI, subspace mapping arises from the observation that Bloch-response trajectories occupy a smooth, low-dimensional manifold and can be well approximated by a finite-dimensional subspace. After applying SVD to a dictionary of signal responses $M^* = X_S'X_T$3, the leading $M^* = X_S'X_T$4 singular vectors $M^* = X_S'X_T$5 provide a basis such that

$M^* = X_S'X_T$6

The reconstruction is then performed entirely in subspace via

$M^* = X_S'X_T$7

while both the dictionary and its Jacobians are projected into the same subspace. Parameter mapping proceeds by reference matching and a local linear inversion using the stored Jacobian, producing continuous-valued estimates of $M^* = X_S'X_T$8, $M^* = X_S'X_T$9, and proton density (Roeloffs et al., 2018).

This MRI formulation combines two distinct subspace mechanisms. First, it uses subspace-constrained reconstruction to compress temporal signal evolution. Second, it uses adaptive tiny dictionaries and manifold projection to reduce the dictionary size by one to two orders of magnitude while maintaining agreement with template matching using heuristic dictionaries (Roeloffs et al., 2018). The detailed exposition states that high-fit accuracy could be achieved with as few as 181 entries, compared to more than 10,000 in a traditional schema, and that four subspace coefficients sufficed in the reported phantom mapping experiments. These are application-specific consequences of representing both data and model in a common signal subspace.

A different intermediary use of subspace mapping appears in robotic teleoperation. There, the goal is to bridge the kinematic mismatch between a human hand and a non-anthropomorphic robot hand through a continuous teleoperation subspace. The shared subspace is three-dimensional, with coordinates corresponding to size $A \in {\mathbb C}^{m \times n}$0, spread $A \in {\mathbb C}^{m \times n}$1, and curl $A \in {\mathbb C}^{m \times n}$2. For a hand with joint vector $A \in {\mathbb C}^{m \times n}$3, origin pose $A \in {\mathbb C}^{m \times n}$4, projection matrix $A \in {\mathbb C}^{m \times n}$5, and scaling factor $A \in {\mathbb C}^{m \times n}$6, the projection rules are

$A \in {\mathbb C}^{m \times n}$7

and

$A \in {\mathbb C}^{m \times n}$8

The full master-to-slave mapping composes these hand-specific transforms through the shared teleoperation subspace (Meeker et al., 2019).

Two procedures are described for constructing the teleoperation subspace. The empirical mapping depends on a designer identifying basis motions from the hand’s kinematics, while the algorithmic mapping is kinematically independent and builds the subspace from grasp data on canonical objects, using Random Sample Consensus and orthogonalization to recover three basis vectors and an origin (Meeker et al., 2019). Experimental validation with ten novice users and two kinematically distinct hands showed that both mappings were intuitive enough to enable effective real-time teleoperation. In the earlier experimental report, novice users completed pick-and-place tasks significantly faster with subspace mapping than with fingertip mapping or joint mapping, with average times of $A \in {\mathbb C}^{m \times n}$9 seconds for subspace mapping, $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$0 seconds for fingertip mapping, and $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$1 seconds for joint mapping (Meeker et al., 2018).

5. Robust concept erasure in diffusion models

In text-to-image diffusion models, SuMa is introduced as a method for concept erasure that aims to achieve both robustness and effectiveness for narrow concepts such as copyrighted characters or celebrities. The core observation is that a concept is spread across a subspace in token-embedding space rather than contained in a single token. On that basis, the method first derives a target subspace representing the concept to be erased and then neutralizes it by mapping it to a reference subspace that minimizes the distance between the two (Nguyen et al., 6 Sep 2025).

The workflow has two stages. In subspace construction, iterative textual inversion is run several times to obtain learned tokens that can reconstruct the target concept in the modified model; these tokens span the target subspace $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$2. For each textual inversion run, “early” tokens from initial optimization steps are also extracted; these tend to encode a more general or related concept and define the reference subspace $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$3. For the $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$4-th U-Net cross-attention layer, the transformed target and reference tokens are written as

$\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$5

In subspace elimination, the method fine-tunes the cross-attention weights so that the target representations are projected onto the reference subspace. With projection matrix

$\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$6

the projection loss is

$\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$7

and a regularization term preserves the reference-subspace structure. The combined subspace loss is

$\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$8

which is added to a Concept Ablation loss to maintain denoising behavior (Nguyen et al., 6 Sep 2025).

The empirical scope is broad: subclass erasure, celebrity erasure, artistic style erasure, and instance erasure are all reported. Evaluation uses Attack Success Rate for robustness and FID and CLIP score for effectiveness. The reported result is that SuMa achieves image quality comparable to approaches focused on effectiveness while also yielding results on par with methods targeting completeness. The detailed summary further states that SuMa achieves low ASR, specifically below $\sum_{i=1}^p \sum_{j=1}^q B_iX_{ij}C_j,$9 for all categories under attack, with FID similar to effectiveness-oriented concept erasure methods and much better than STEREO, and that it is the first method to achieve both robustness and utility on instance and identity erasure (Nguyen et al., 6 Sep 2025). In this usage, subspace mapping is neither a geometric alignment nor a reduced-order model; it is a targeted neutralization of semantic directions in a learned generative model.

6. Quantum, entanglement, and many-body formulations

In quantum linear algebra, subspace mapping takes the form of quantum subspace states. For a matrix $B_i \in {\mathbb C}^{m \times g_i}$0 with orthonormal columns, the associated state is

$B_i \in {\mathbb C}^{m \times g_i}$1

which encodes a $B_i \in {\mathbb C}^{m \times g_i}$2-dimensional subspace of $B_i \in {\mathbb C}^{m \times g_i}$3. This representation underlies a quantum determinant sampling algorithm, a singular value estimation algorithm for compound matrices, and a reduction in circuit depth for quantum topological data analysis. The determinant sampling routine prepares the subspace state and samples from the determinant distribution in $B_i \in {\mathbb C}^{m \times g_i}$4 gates with circuit depth $B_i \in {\mathbb C}^{m \times g_i}$5, while the TDA construction reduces the relevant circuit depth from $B_i \in {\mathbb C}^{m \times g_i}$6 to $B_i \in {\mathbb C}^{m \times g_i}$7 (Kerenidis et al., 2022). Here the mapped object is a subspace itself, represented as a quantum state rather than as a basis matrix.

A distinct quantum-information use appears in entanglement theory. For even $B_i \in {\mathbb C}^{m \times g_i}$8, a mapping

$B_i \in {\mathbb C}^{m \times g_i}$9

takes an $C_j \in {\mathbb C}^{h_j \times n}$0-qubit symmetric state to a bipartite symmetric state of higher local dimension. On Dicke states,

$C_j \in {\mathbb C}^{h_j \times n}$1

The mapping preserves inner products and separability, so entanglement of the image implies entanglement of the original state. It is then used to bound the symmetric tensor rank and the geometric measure of entanglement, and to identify an orthogonal complementary subspace in which all pure states are entangled (Marconi et al., 2 Apr 2025). This is a case where subspace mapping functions as a complexity-reducing change of problem class: multipartite symmetric entanglement is studied through a bipartite symmetric surrogate.

In boson expansion theory, the phrase “subspace mapping” appears in a more critical, corrective role. The norm operator method is used to derive subspace mapping properly from whole space mapping, and the derivation requires renormalization of phonon contributions that are not adopted as boson excitations in the subspace mapping. The paper states that conventional boson expansion theories ignore these contributions, and that simple truncation is correct only when the non-adopted phonon excitations make no contribution at all. It also argues that the Park operator, previously considered effective only in whole space mapping, is also effective in subspace mapping (Taniguchi, 19 Nov 2025). This usage is notable because it directly addresses a controversy: whether subspace restriction can be justified merely by discarding non-adopted modes. The answer given is conditional rather than universal.

Taken together, these quantum and many-body formulations show that SuMa can refer to basis-dependent projection, state encoding, separability-preserving embedding, or renormalized restriction. The shared structure remains recognizable: a high-dimensional object is replaced by a carefully chosen subspace representation or by a map between subspaces, and the validity of that replacement depends on what invariants the construction is designed to preserve.