---
title: Expanded Sufficiently Scattered Condition
url: https://www.emergentmind.com/topics/expanded-sufficiently-scattered-condition
type: topic
---

# Expanded Sufficiently Scattered Condition

The expanded sufficiently scattered condition (p-SSC) is a quantitative structural property introduced to precisely characterize the geometric spread of data in minimum-volume nonnegative matrix factorization (min-vol NMF). It has emerged as a central tool for establishing the noise robustness of min-vol NMF, bridging the gap between the classical notions of separability and the standard sufficiently scattered condition (SSC). The p-SSC formulates a new rigorously defined measure of how well the rows of the factor matrix fill out the latent simplex, parameterized by $p\in [1,\sqrt{r-1})$, and enables sharp finite-noise recovery guarantees for min-vol NMF in the presence of perturbations.

## 1. Model Formulation and Latent Simplex Structure

Consider a data matrix $X = [x_1, \dots, x_n] \in \mathbb{R}^{m \times n}$ modeled as
\[
X \approx W H^\top,\qquad W\in\mathbb{R}^{m\times r},\quad H\in\mathbb{R}^{n\times r},\quad r \ll \min\{m,n\},
\]
with a simplex-structured constraint
\[
H e = e,\quad H \ge 0, \quad x_i = W H(i,:)^\top \quad\Longrightarrow\quad x_i\in\operatorname{Conv}\{W(:,1),\dots,W(:,r)\},
\]
where $e$ is the all-ones vector in $\mathbb{R}^{r}$. Often, data columns are normalized so that
\[
\Delta^r = \left\{h \in \mathbb{R}_{\geq 0}^r \mid e^\top h = 1\right\}
\]
and thus $\operatorname{Conv}(X) \subseteq \operatorname{Conv}(W)$. This normalization ensures every data column lies in the standard probability simplex, rendering the geometric analysis of the factorization meaningful in terms of volume and scatter.

## 2. Definition of the Expanded Sufficiently Scattered Condition

For $p\geq1$, define the cone
\[
C_p = \left\{x\in\mathbb{R}_{\ge 0}^r \mid e^\top x \geq p\|x\|\right\}.
\]
On the affine hyperplane $\{x \mid e^\top x = 1\}$, this corresponds to the set
\[
Q_p = \left\{x = \frac{e}{r} + w\;\Big|\; e^\top w = 0,\; \|w\|^2 \leq \frac{1}{p^2} - \frac{1}{r}\right\} \subseteq \Delta^r,
\]
with $C_p = \operatorname{Cone}(Q_p)$.

A matrix $H\in\mathbb{R}^{n\times r}_{\geq 0}$ is $p$-SSC if and only if
\[
C_p \subseteq \operatorname{Cone}(H^\top)\qquad\text{or, equivalently,}\qquad Q_p \subseteq \operatorname{Conv}\{H(1,:),\dots,H(n,:)\}.
\]
The $p$-SSC thus demands that the convex hull of the rows of $H$ contains a Euclidean ball of radius determined by $p$, centered at the barycenter of the simplex.

**Key regimes:**

| $p$ value | Condition                                      | Interpretation                                 |
|-----------|------------------------------------------------|------------------------------------------------|
| $p=1$     | $C_1 = \mathbb{R}_{\geq 0}^r$                  | Separability                                   |
| $p=\sqrt{r-1}$ | Standard sufficiently scattered condition (SSC) | Inscribed sphere in $\operatorname{Conv}(H^\top)$ |
| $1<p<\sqrt{r-1}$ | Expanded SSC (stronger than SSC)              | Excludes near-flat configurations              |

## 3. Geometric Interpretation and Comparison to Classical SSC

Classical SSC ($p=\sqrt{r-1}$) requires the largest possible inscribed sphere of $\Delta^r$ to be contained in $\operatorname{Conv}(H^\top)$, forcing the $H(i,:)$ to occupy positions relatively far from simplex faces. In contrast, the p-SSC for $1\leq p<\sqrt{r-1}$ tightens this by requiring that a **larger** sphere $Q_p \subsetneq Q_{\sqrt{r-1}}$ is contained. Geometrically, p-SSC rules out degenerate configurations where data points are nearly affine-dependent or clustered near faces of the simplex, which degrade identifiability.

As $p\to\sqrt{r-1}$, the region allowed by the condition shrinks, and the tolerance for noise diminishes. As $p\to1$, one approaches separability, admitting maximal robustness.

## 4. Robustness Guarantees for Minimum-Volume NMF under p-SSC

Given the approximate min-vol NMF formulation:
\[
\min_{W,H} \;\det(W^\top W) \quad
\text{s.t.}\;
\|X-WH^\top\|_{1,2} \leq \varepsilon,\;
H e = e,\;
H \geq 0,
\]
where $\|A\|_{1,2} = \max_j \|A(:,j)\|$, and the generative model
\[
X = W^{\#}(H^{\#})^\top + N^{\#},\qquad H^{\#} \text{ is } p\text{-SSC},\quad \operatorname{rank}(W^{\#}) = r,\quad \|N^{\#}\|_{1,2} \leq \varepsilon,
\]
the main recovery theorem asserts—letting $q = \sqrt{r - p^2}$—that there exist absolute constants $C_1, C_2 > 0$ such that if
\[
\varepsilon \leq C_1 \frac{\sigma_r(W^{\#})}{r^{9/2} \frac{q^2}{p^2} (\min\{q, \sqrt{2}\} - 1)^2},
\]
then for any optimizer $(W^*, H^*)$, the factor recovery error obeys
\[
\min_{\Pi\in\mathcal{P}_r} \|W^{\#} - W^*\Pi\|_{1,2}
      \leq
      C_2 \|W^{\#}\|
      \sqrt{
        \frac{\varepsilon}
             {\min\{q^2-1,\,1\}
              r^{7/2}\,\sigma_r(W^{\#})
              \frac{p^2}{q^2}
        }
      },
\]
where $\Pi$ runs over all $r\times r$ permutation matrices, and $\sigma_r(W^{\#})$ is the smallest nonzero singular value of $W^{\#}$. In the near-separable regime $p\to1$, this simplifies to
\[
\min_{\Pi} \|W^{\#}-W^*\Pi\|_{1,2}
      = O(\|W^{\#}\|)\left(
        \frac{r\sqrt r}{\sigma_r(W^{\#})} + r(p-1)
      \right).
\]

The bounds underscore that robustness depends critically on $p$, $q$, the conditioning of $W^{\#}$, and the noise bound $\varepsilon$. In particular, tightness of the bound on $\varepsilon$ diminishes rapidly as $p$ approaches $\sqrt{r-1}$ (i.e., standard SSC), and is strongest in the near-separable regime ($p \to 1$). The result establishes—quantitatively and for the first time—that the expanded sufficiently scattered condition yields provable stability for min-vol NMF under explicit noise models.

## 5. Principal Geometric Lemmas

A suite of geometric lemmas underpins the analysis of p-SSC:

- **Dual cone:** For a cone $K\subseteq\mathbb{R}^r$, the dual is $K^* = \{y \mid x^\top y \geq 0,\,\forall x\in K\}$.
- **Geometry of $Q_p$ and associated cones:** $C_p\cap\{e^\top x=1\} = Q_p$, and $C_p^*=S_q + \mathbb{R}^r_{\geq 0}$, with $S_q = \{x \mid e^\top x\geq q\|x\|\}$ and $q=\sqrt{r-p^2}$.
- **Auxiliary simplex and necessary convex hull:** Any $p$-SSC $H$ must satisfy $\operatorname{Conv}(H_p^\top)\subseteq\operatorname{Conv}(H^\top)\subseteq\Delta^r$ for $H_p^\top = \alpha_pE+(1-r\alpha_p)I$, with $\alpha_p = \frac{1}{r}(1-\frac{1}{\sqrt{r-1}q/p})$.
- **Lower bound on determinant of linking map:** There exists $R\in\mathbb{R}^{r\times r}$ with $W^{\#}=W^* R+(\text{small})$ and $\det(R)^2 \geq 1 - O(r^2/\sigma_r(W^{\#})\cdot p/q)$.
- **Approximate orthogonality:** Writing $r_i^\top$ for the $i$th row of $R$, $e^\top r_i\approx1$, $\|r_i\|\approx1$, $\|r_i-r_j\|\approx1$ up to $O(\sqrt{\varepsilon}\,p^2/q^2\,r^{7/2}/\sigma_r(W^{\#}))$.
- **Geometric localization:** Each $r_i$ lies in a small "ice-cream" sector around a canonical basis vector $e_k$, with angular radius $O(\sqrt{\frac{\varepsilon}{\min\{q^2-1,1\}\,r^{7/2}\,\sigma_r(W^{\#})\,p^2/q^2}})$; distinct rows cannot coincide at the same vertex.

These elements facilitate the structural identification and stability analysis under the p-SSC framework.

## 6. Parametric Dependencies and Regime Analysis

The roles of the principal parameters are as follows:

- $\varepsilon$ (noise magnitude): Smaller $\varepsilon$ is required for successful recovery as $p$ increases; robustness is sharply diminished for larger $p$, i.e., as the sphere inscribed in the simplex shrinks.
- $p\in [1, \sqrt{r-1})$: Encodes the "well spread" property of ground-truth rows of $H^{\#}$; regimes interpolate between separability ($p\to1$), which is maximally robust, and the classical SSC ($p\to\sqrt{r-1}$), which admits no fixed noise tolerance.
- $q=\sqrt{r-p^2}$: The “dual slack” parameter appearing in bounds; smaller $q$ (i.e., $p$ closer to $\sqrt{r-1}$) both tightens the allowable $\varepsilon$ and worsens the final error.
- $\sigma_r(W^{\#})$, $\|W^{\#}\|$: Reflect the conditioning of the true basis matrix; poor conditioning directly increases sensitivity to noise.

A plausible implication is that, for practical applications requiring noise robustness in min-vol NMF, it is advantageous for data to admit representations with $p$ as close to $1$ as possible, and for $W^{\#}$ to be well-conditioned.

## 7. Summary and Implications

The expanded sufficiently scattered condition (p-SSC) provides a continuum of structural regimes interpolating between separability and the classical sufficiently scattered condition. By enforcing a suitably large inscribed sphere in the latent simplex, p-SSC enables explicit, quantitative guarantees for the identifiability and robustness of minimum-volume NMF under bounded perturbations. The derived bounds demonstrate that when data are sufficiently well-scattered (with smaller $p$), min-vol NMF becomes provably stable, while for data satisfying only the classical SSC, robustness rapidly deteriorates. The introduction of p-SSC thus offers a principled geometric criterion for algorithmic design and analysis in applications of NMF subject to noise, enhancing the theoretical understanding and guiding practical data preconditioning.

Source: https://www.emergentmind.com/topics/expanded-sufficiently-scattered-condition