Strong Singular Value Property

• Strong Singular Value Property is a condition ensuring that only the zero matrix satisfies the symmetry and zero-pattern constraints, thereby spanning all perturbation directions.
• It underpins inverse singular-value problems by enabling superpattern and bifurcation theorems that allow singular value lists to be achieved despite strict pattern restrictions.
• The concept extends to certifiable sparse singular value analyses and random-matrix decoupling, impacting robust statistics and spectral gap estimation.

Strong Singular Value Property denotes several distinct notions in recent singular-value literature. In one formal usage, introduced for real matrices, it is a transversality condition governing infinitesimal orthogonal left-right actions together with perturbations on the zero positions of a prescribed pattern; this version is used to analyze which lists of nonnegative real numbers occur as the singular values of a matrix with a prescribed zero-nonzero pattern (Cheung et al., 11 Jul 2025). In adjacent lines of work, related expositions use the same acronym for a Sum-of-Squares certifiability condition for sparse singular values of random rectangular matrices and for a quantitative decoupling phenomenon for least singular values of random symmetric matrices (Diakonikolas et al., 2024, Han, 22 Apr 2025). The expression is therefore context-dependent. By contrast, the C$^*$-algebra paper "Singular value functions for C$^*$-algebras" studies singular value functions $s(a):K_0(A)^+\to \mathbb R_+$ and explicitly does not introduce any notion called the Strong Singular Value Property (Fujitsu, 17 May 2026).

1. Scope and nomenclature

The most precise matrix-theoretic definition appears in "The Strong Singular Value Property for Matrices," where the property is attached to a real $m\times n$ matrix $A$ with $m\le n$ and singular values $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$ (Cheung et al., 11 Jul 2025). In that setting, SSVP is a local linear-algebraic condition tied to zero-pattern geometry and inverse singular-value realizability.

A separate usage, presented in the exposition associated with "SoS Certificates for Sparse Singular Values and Their Applications: Robust Statistics, Subspace Distortion, and More," concerns an $\eta$-sparse singular value

$$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$

and asks for efficiently verifiable low-degree SoS certificates that $\sigma_{\max}^\eta(M)\le B$ (Diakonikolas et al., 2024). Here the emphasis is algorithmic certification rather than pattern rigidity.

A third usage, summarized for "Repeated singular values of a random symmetric matrix and decoupled singular value estimates," treats SSVP as a quantitative statement that small least-singular-value events at two separated shifts $^*$0 essentially factorize up to a $^*$1 term and exponentially small error (Han, 22 Apr 2025). This is a random-matrix decoupling phenomenon rather than a property of a fixed deterministic matrix pattern.

Usage	Object	Core content
Matrix SSVP	Real $^*$2 matrix	Only $^*$3 satisfying symmetry and zero-support constraints is $^*$4
Strong sparse singular value property	Random rectangular matrix	SoS-certifiable upper bound on $^*$5
Random-matrix SSVP	Wigner matrix with two shifts	Joint least-singular-value tail decoupling

The nearby C$^*$6-algebra literature on singular value functions is relevant only terminologically. The paper (Fujitsu, 17 May 2026) introduces singular value functions for C$^*$7-algebras and develops their basic properties, but does not define or mention SSVP.

2. Matrix-theoretic definition

For a real $^*$8 matrix $^*$9, let $s(a):K_0(A)^+\to \mathbb R_+$0 denote the Schur product. The matrix $s(a):K_0(A)^+\to \mathbb R_+$1 has the Strong Singular Value Property if the only $s(a):K_0(A)^+\to \mathbb R_+$2 satisfying

1. $s(a):K_0(A)^+\to \mathbb R_+$3 is symmetric,
2. $s(a):K_0(A)^+\to \mathbb R_+$4 is symmetric,
3. $s(a):K_0(A)^+\to \mathbb R_+$5,

is $s(a):K_0(A)^+\to \mathbb R_+$6 (Cheung et al., 11 Jul 2025).

The same paper gives an equivalent subspace formulation. Define

$s(a):K_0(A)^+\to \mathbb R_+$7

Then

$s(a):K_0(A)^+\to \mathbb R_+$8

This identifies SSVP as a spanning condition in the ambient matrix space (Cheung et al., 11 Jul 2025).

The geometric interpretation given there is that tangent directions arising from left- and right-infinitesimal orthogonal conjugations, together with arbitrary changes in the zero positions, span the full ambient space. In the same source, the derivative at $s(a):K_0(A)^+\to \mathbb R_+$9 of the map

$m\times n$0

has image exactly $m\times n$1, so SSVP augments the local orbit directions by zero-position perturbations until full surjectivity is reached (Cheung et al., 11 Jul 2025). This makes SSVP a differential local-surjectivity condition for singular-value realizability under pattern constraints.

3. Structural criteria, characterizations, and examples

Several basic criteria delimit when SSVP can or cannot occur. A first necessary condition is full term-rank: if $m\times n$2 has SSVP, then $m\times n$3 has term-rank $m\times n$4, meaning that there is a choice of $m\times n$5 ones in the pattern with no two in the same row or column (Cheung et al., 11 Jul 2025). This is a combinatorial prerequisite.

The paper gives a complete characterization for diagonal matrices. A diagonal matrix $m\times n$6 has SSVP if and only if $m\times n$7 for all $m\times n$8, and $m\times n$9 whenever $A$0 (Cheung et al., 11 Jul 2025). Thus $A$1 has SSVP, while $A$2 fails because $A$3.

For direct sums, if $A$4 is $A$5 and $A$6 is $A$7 with $A$8, then

$A$9

if and only if $m\le n$0, $m\le n$1, both $m\le n$2 and $m\le n$3 have SSVP, $m\le n$4 and $m\le n$5 share no common nonzero singular value, and either both have full row-rank or one is square and invertible (Cheung et al., 11 Jul 2025). This criterion shows that SSVP is sensitive both to pattern and to spectral collisions.

The 2-by-$m\le n$6 case admits a normal-form analysis. Up to row and column permutations and sign-changes, any $m\le n$7 matrix with no zero column can be taken to the block form

$m\le n$8

after which a direct linear-algebraic argument determines exactly when SSVP holds; the summary notes that SSVP fails whenever there is a zero row or too much symmetry between the two supports (Cheung et al., 11 Jul 2025).

Concrete examples sharpen these criteria. Nowhere-zero matrices trivially have SSVP because $m\le n$9. Any full row-orthonormal $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$0 matrix $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$1 with $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$2 has SSVP. A matrix with a zero row fails SSVP, and the failure can be witnessed by a nonzero rank-one $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$3 supported in that row and annihilated by $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$4 (Cheung et al., 11 Jul 2025).

The paper also introduces a finite-dimensional verification test. An explicit verification matrix $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$5 encodes the linear system given by the three defining conditions, with columns indexed by zero positions of $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$6 and rows coming from the coefficients of the maps $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$7 and $$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$8. Then

$$\Sigma(A)=\{\sigma_1(A)\ge \cdots \ge \sigma_m(A)\ge 0\}$$9

This reduces verification to a linear algebra computation (Cheung et al., 11 Jul 2025).

4. Inverse singular-value problems and superpatterns

The principal application of matrix SSVP is the inverse singular-value problem for zero-nonzero patterns. The paper states a Superpattern Theorem: if $\eta$0 has SSVP and pattern $\eta$1, then every superpattern of $\eta$2 admits a realization of the same $\eta$3 (Cheung et al., 11 Jul 2025). The proof sets up

$\eta$4

on skew-symmetric perturbations together with zero-position directions and applies the Inverse Function Theorem to the onto derivative at $\eta$5.

A complementary Bifurcation Theorem states that if $\eta$6 has SSVP and pattern $\eta$7, then any nearby $\eta$8 in Euclidean norm can be realized by some $\eta$9 with pattern $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$0 and $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$1 (Cheung et al., 11 Jul 2025). Equivalently, SSVP allows one to move the singular-value list arbitrarily while holding the pattern.

The Matrix Liberation Theorem extends the framework beyond matrices that themselves satisfy SSVP. Even if $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$2 fails SSVP, one can sometimes free a new zero position to a nonzero one without losing singular-value realizability, provided a certain tangential-span condition holds (Cheung et al., 11 Jul 2025). This again rests on an extension of the Inverse Function Theorem.

These results situate SSVP as the singular-value analogue of local rigidity removal. The same source states that SSVP is the singular-value analogue of the Strong Spectral Property from the symmetric-matrix literature, and notes as an open direction a complete combinatorial characterization of which patterns admit at least one SSVP matrix (Cheung et al., 11 Jul 2025). The summary further suggests that, beyond term-rank, one must understand forbidden substructures in the associated bigraphs. A plausible implication is that SSVP organizes inverse singular-value theory around transversality rather than around ad hoc pattern-specific constructions.

5. Sparse singular values and Sum-of-Squares certificates

In the exposition associated with (Diakonikolas et al., 2024), the phrase "strong (sparse) singular value property" refers to certificate-based control of sparse singular values for random rectangular matrices. For $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$3 and $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$4,

$$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$5

An equivalent formulation introduces Boolean selectors $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$6 with $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$7, and interprets the problem as selecting at most $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$8 rows of $$\sigma_{\max}^\eta(M)=\max_{\substack{u\in\mathbb R^n\ \|u\|_2=1,\;\|u\|_0\le \eta n}}\|Mu\|_2$$9 (Diakonikolas et al., 2024).

A certificate that $\sigma_{\max}^\eta(M)\le B$0 is any efficiently verifiable proof, in particular a low-degree SoS proof, of the polynomial implication

$\sigma_{\max}^\eta(M)\le B$1

or equivalently

$\sigma_{\max}^\eta(M)\le B$2

under the same constraints (Diakonikolas et al., 2024). The primal formulation uses a degree-$\sigma_{\max}^\eta(M)\le B$3 pseudo-expectation operator satisfying these polynomial constraints, and SoS duality equates this with the existence of a low-degree SoS proof.

For Gaussian $\sigma_{\max}^\eta(M)\le B$4, the exposition states an informal theorem: for any $\sigma_{\max}^\eta(M)\le B$5, there is an SoS proof of degree $\sigma_{\max}^\eta(M)\le B$6 certifying

$\sigma_{\max}^\eta(M)\le B$7

with overwhelming probability, provided $\sigma_{\max}^\eta(M)\le B$8 (Diakonikolas et al., 2024). The proof strategy combines a Schatten-$\sigma_{\max}^\eta(M)\le B$9 relaxation, an expansion of $^*$00, graph-polynomial grouping, Efron-Stein decomposition, graph-matrix spectral bounds, and an induction on $^*$01.

The same framework underlies applications listed in the source: robust covariance estimation, covariance-aware mean estimation, Euclidean mean estimation, certification of $^*$02 distortion of random subspaces, sparse principal component analysis, and certification of the $^*$03 norm of a random matrix (Diakonikolas et al., 2024). This suggests that, in this line of work, the phrase functions as a shorthand for a certifiable sparse-operator-norm phenomenon rather than for the zero-pattern transversality condition of (Cheung et al., 11 Jul 2025).

6. Random matrices, least singular values, and neighboring notions

The summary attached to (Han, 22 Apr 2025) uses SSVP for a two-point small-singular-value estimate for random symmetric matrices. If $^*$04 is an $^*$05 Wigner matrix with real symmetric structure and $^*$06 lie in the bulk interval $^*$07, then for Bernoulli $^*$08 entries, or more generally for subgaussian entries with a finite log-Sobolev constant, the theorem states that when $^*$09,

$^*$10

For general subgaussian laws and mesoscopic separation $^*$11, the analogous bound is

$^*$12

These estimates are described as showing that extreme behaviors of the least singular value at two locations can essentially be decoupled when the shifts are separated (Han, 22 Apr 2025).

A corollary in the same summary states that for

$^*$13

the probability that there is a repeated singular value in $^*$14 is at most $^*$15, and the minimal gap among singular values in $^*$16 is at least constant$^*$17 with high probability (Han, 22 Apr 2025). The techniques listed include the local semicircle law, Hanson-Wright inequality, Talagrand concentration, inverse Littlewood-Offord bounds, a zero-out matrix $^*$18, Esseen's lemma in two dimensions, and a bootstrap iteration.

This random-matrix usage is conceptually distinct from the deterministic matrix-pattern definition and from the SoS sparse-certification framework. It concerns repulsion and decoupling of least singular values across spectral shifts, not local surjectivity of a singular-value map. The contrast is sharpened by the C$^*$19-algebra case: "Singular value functions for C$^*$20-algebras" develops singular value functions $^*$21, including non-negativity, monotonicity, subadditivity, Ky-Fan-type inequalities, continuity under real-rank-zero hypotheses, realization theorems, and examples for AF algebras and compact operators, but nowhere defines or studies a Strong Singular Value Property (Fujitsu, 17 May 2026). The term therefore does not designate a single established concept across all singular-value literatures.