MAS-SVD: Magnitude-Aware Spherically Normalized SVD

• The paper introduces a hybrid approach that combines spherical normalization for robust directional recovery with a magnitude-aware mechanism that restores lost scale information.
• It employs two reduced-rank SVD calls with discrete pairing and weighted-median estimation to ensure resilient singular vector extraction under norm distortions.
• The method bridges classical SVD and spherical matrix factorization, offering improved robustness against outliers while preserving both directional and magnitude details.

Searching arXiv for the cited papers and closely related terminology to ground the article. Magnitude-Aware Spherically Normalized SVD (MAS-SVD) is best understood, in the present arXiv literature, not as a separately formalized decomposition with a single canonical definition, but as a conceptual extension of two closely related lines of work: spherical normalization for robust singular-vector recovery and spherical constraints in low-rank factorization. In that reconstruction, the “spherically normalized” component refers to estimating latent or singular directions after enforcing unit-norm or equal-norm geometry, while the “magnitude-aware” component refers to restoring scale information that plain spherical normalization suppresses. The most direct foundations are "Spherically Normalized SVD" (Han et al., 2024) and "Spherical Matrix Factorization" (Liu, 2021).

1. Terminological status and scope

The expression “Magnitude-Aware Spherically Normalized SVD” does not appear as the title or explicit method name in the cited papers. The relevant literature instead provides a robust SVD method called Spherically Normalized SVD (SpSVD) and, separately, a constrained factorization framework in which latent codes are constrained to a sphere. Accordingly, MAS-SVD is most precisely treated as an inferred research construct: an SVD-shaped model that retains the directional robustness or angular geometry of spherical normalization while avoiding the complete loss of radial information that ordinary normalization induces (Han et al., 2024).

In SpSVD, the basic rank-$R$ target is

$$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$

Its defining idea is to estimate candidate right singular vectors from a row-normalized matrix and candidate left singular vectors from a column-normalized matrix, then select rank-one terms on the original scale using an elementwise $L_1$-type criterion. The paper is explicit that this is not iterative alternating normalization, not RPCA, and not an $L_1$-SVD solved by repeated optimization over all singular vectors. It is a two-SVD construction followed by discrete pairing and weighted-median scale estimation.

The spherical matrix factorization paper occupies a different but complementary position. It studies

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$

with hard equal-norm constraints on latent columns $v_j$. That framework is not an SVD derivation, but it formalizes the same geometric principle that motivates a magnitude-aware spherical SVD: Euclidean reconstruction can be combined with latent angular structure by constraining codes to a sphere.

2. Spherical normalization and angular geometry

The central geometric mechanism is the replacement of unconstrained latent or observed vectors by normalized ones, so that Euclidean separation becomes a monotone function of angle on a fixed-radius sphere. In spherical matrix factorization, the master problem is

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$

with feasible sets that include

$$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$

and

$\mathcal V_1:=\{V\mid \|V(:,j)\|=l,\ \forall j\},$

together with nonnegative and sparse variants $\mathcal V_2,\mathcal V_3,\mathcal V_4$. There is no explicit cosine-similarity term added to the objective; angular structure is introduced entirely through the hard constraints on $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$0 (Liu, 2021).

The paper’s key identity is that, if $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$1, then

$$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$2

Thus, on a fixed-radius sphere, Euclidean distance and angle induce equivalent orderings. This is the precise sense in which spherical normalization “unifies” Euclidean and angular geometry: reconstruction remains Euclidean, while representation geometry becomes directional.

SpSVD applies the same logic directly to the data matrix rather than to latent codes. If $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$3, the row-normalized matrix is

$$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$4

and if $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$5, the column-normalized matrix is

$$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$6

Row normalization preserves row direction and discards row magnitude; column normalization preserves column direction and discards column magnitude. This directionalization is exactly what makes the method insensitive to arbitrarily large row or column norms.

A common misconception is that spherical normalization introduces a mixed Euclidean–cosine loss. In the cited formulations, it does not. The Euclidean or $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$7-type reconstruction criterion is kept separate from the normalization mechanism. The angular effect arises from feasible-set geometry or preprocessing, not from adding explicit cosine penalties.

3. Algorithmic backbone of spherically normalized SVD

SpSVD has a concrete procedural definition. First, compute $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$8 and $$X = UDV^T = \sum_{r=1}^{n\land p} d_r\, u_r v_r^T, \qquad \widehat{X}_R^{\,s} := \sum_{r=1}^{R} d_r^{\,s}\, u_r^{\,s} (v_r^{\,s})^T.$$9. Next, apply a standard rank-$L_1$0 SVD algorithm to each: $L_1$1 The paper emphasizes that one should not naively pair

$L_1$2

because the two normalizations can induce different orderings (Han et al., 2024).

The first component is selected by

$L_1$3

where

$L_1$4

For $L_1$5, one forms the residual

$L_1$6

removes already used candidates,

$L_1$7

and solves

$L_1$8

For fixed $L_1$9 and $L_1$0, the scalar $L_1$1 is obtained as a weighted median by solving

$L_1$2

If $L_1$3, the sign is flipped so that singular values are nonnegative: $L_1$4

A defining implementation property is that the method uses only two reduced-rank SVD calls. Subsequent work consists of discrete pairing, weighted-median scalar estimation, and sequential deflation. The paper states total computational complexity

$L_1$5

This places the method close to standard rank-$L_1$6 SVD when $L_1$7 is small, while shifting robustness into preprocessing and pairing rather than iterative factor optimization.

4. Magnitude awareness: what spherical normalization removes and how scale returns

The phrase “magnitude-aware” is most meaningful in relation to what ordinary spherical normalization discards. In SpSVD, the transformations

$L_1$8

force all row norms or column norms to 1. The normalized SVDs therefore no longer encode the original singular values of $L_1$9, nor the energy distribution across rows and columns. Scale is reintroduced only afterward, through the robust scalar fit $$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$0 on the original residual matrix (Han et al., 2024).

Spherical matrix factorization exhibits the same tension in a latent-variable setting. The latent codes satisfy

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$1

so sample-specific latent magnitudes are not preserved. The model retains only the direction of each latent code and a single shared global scale $$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$2. After solving unit-norm directional subproblems, the paper updates the common radius by

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$3

This is a clean separation of direction and scale, but only at the level of a global scalar (Liu, 2021).

From a MAS-SVD perspective, this identifies the exact structural limitation of plain spherical normalization. It is direction-aware but not fully magnitude-aware. The factorization paper itself notes the decomposition

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$4

and directly suggests the more expressive form

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$5

or the introduction of a diagonal matrix $$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$6. This suggests a natural MAS-SVD interpretation: preserve the spherical directional variables, but replace a single shared scale by samplewise, componentwise, or singular-value-like magnitude structure.

A related misconception is that SpSVD already solves magnitude preservation because it estimates singular values after normalization. That is only partially correct. The scalar $$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$7 restores one coefficient per rank-one term, but the candidate left and right singular vectors are still extracted from fully normalized data. Magnitude therefore does not influence the directional candidate sets in the basic method.

5. Relation to SVD, PCA, and constrained spherical factorization

The conceptual bridge from standard SVD/PCA to a magnitude-aware spherical SVD is explicit in the factorization framework. The paper starts from PCA in the form

$$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$8

for which the unconstrained optimum is $$X \approx UV,\qquad X\in\mathbb{R}^{m\times n},\; U\in\mathbb{R}^{m\times r},\; V\in\mathbb{R}^{r\times n},$$9. Once spherical, sparse, or nonnegative constraints are imposed on $v_j$0, this closed-form score update disappears, and the model becomes a constrained analogue of PCA/SVD rather than PCA/SVD itself (Liu, 2021).

The special cases

• $v_j$1: Spherical PCA,
• $v_j$2: Spherical NMF,

show that spherical normalization can be attached to both orthogonal-basis and nonnegative-basis factorizations. The orthogonal case is especially relevant to MAS-SVD because it is closest to an SVD-shaped decomposition. If $v_j$3, one may write

$v_j$4

This resembles a PCA/SVD-like basis with normalized score vectors of common norm $v_j$5, but it differs from standard SVD because there is no diagonal singular value matrix $v_j$6, code magnitudes do not vary by sample, and the left and right factors are constrained asymmetrically.

SpSVD supplies the complementary SVD-side structure. It begins from the singular expansion

$v_j$7

but robustifies candidate singular directions through separate row and column normalization, then chooses pairings by minimizing an $v_j$8 residual on the original matrix. A plausible implication is that MAS-SVD, if formalized, would inherit from SpSVD the two-sided candidate extraction and robust pairing logic, while inheriting from spherical factorization a more explicit norm–direction split.

This is also the most natural route to an SVD-shaped magnitude-aware model: $v_j$9 or, equivalently,

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$0

Those exact models are not given as finished algorithms in the papers, but they are directly suggested by the equations already present there.

6. Optimization, robustness, empirical behavior, and limitations

The spherical matrix factorization problem is nonconvex because of bilinearity and nonconvex constraints. The paper applies PALM to

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$1

with alternating updates

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$2

and, columnwise,

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$3

For orthogonal $$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$4, the subproblem reduces to an orthogonal Procrustes form,

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$5

where

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$6

and $$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$7. For the sphere-only case $$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$8, the column update is simply

$$\min_{U\in\mathcal U,\;V\in\mathcal V} h(U,V)=\|X-UV\|_F^2,$$9

With

$$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$0

the paper states that if

$$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$1

then the objective is monotonically non-increasing, and by PALM theory one obtains global sequence convergence to a critical point with at least sub-linear convergence rate (Liu, 2021).

SpSVD, by contrast, does not rely on iterative convergence over factors. Its theoretical emphasis is robustness. The paper introduces row-wise, column-wise, and block-wise breakdown points for singular-vector and subspace estimators. For standard SVD, it proves that the right singular subspace can break down by corrupting just one row, and the left singular subspace by corrupting just one column. For broad element-wise-loss SVD variants, including $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$2- and Huber-type formulations, the same low-breakdown conclusions persist. SpSVD improves this by establishing nontrivial lower bounds $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$3 and $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$4 for the right and left subspaces after normalization (Han et al., 2024).

The empirical record in the two papers is consistent with these structural claims. In spherical matrix factorization, synthetic data with clusters separated by angle are represented more cleanly when latent codes are constrained to the sphere, and the MNIST digit-3 example with $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$5, $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$6, $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$7, $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$8, and sparsity level $$\mathcal U_1:=\{U\mid U^\top U=I\},\qquad \mathcal U_2:=\{U\mid U\ge 0\},$$9 yields interpretable basis vectors and a spherical latent visualization. In SpSVD, simulations with contaminated low-rank matrices show robust subspace recovery, and the gene-expression example reports average relative errors

$\mathcal V_1:=\{V\mid \|V(:,j)\|=l,\ \forall j\},$0

with

$\mathcal V_1:=\{V\mid \|V(:,j)\|=l,\ \forall j\},$1

The limitations are equally clear. SpSVD is robust because it discards row and column magnitudes before directional estimation; the factorization framework is angle-sensitive because it enforces equal latent norms. Neither method is magnitude-aware in the strong sense of preserving samplewise or componentwise norm information throughout the directional estimation stage. This suggests that the most defensible interpretation of MAS-SVD is as a partially realized research direction: combine spherical normalization’s protection against norm-dominated outliers with an explicit mechanism for retaining or reintroducing radial structure, whether through samplewise radii, componentwise scales, or a diagonal singular-value-like parameterization.