Simplex-Structured Matrix Factorization

• SSMF is a matrix factorization model that decomposes a data matrix into a basis and coefficient matrix with simplex (convex combination) constraints.
• It provides both an algebraic formulation and a geometric interpretation through convex hulls, simplices, and volume criteria, enabling clear identification of latent sources.
• Recent advancements include robust volume minimization, bounded and online variants, and multilayer approaches to handle noise, scalability, and endmember variability.

Searching arXiv for the cited SSMF papers and closely related work to ground the article in the current literature. Simplex-Structured Matrix Factorization (SSMF) denotes a class of matrix factorization models in which a data matrix is approximated as $X \approx W H$ and each column of the coefficient matrix $H$ lies in the probability simplex. Under this constraint, every data column belongs to the convex hull of the columns of $W$, so SSMF admits both an algebraic formulation and a geometric one in terms of simplices, convex hulls, facets, and volume criteria. The framework appears in hyperspectral unmixing, chemometrics, topic modeling, document clustering, and related latent-source problems, and has developed along several lines: minimum-volume formulations, inscribed-ellipsoid methods, dual volume maximization, facet-identification approaches, probabilistic interpretations based on determinant regularization, online updates, bounded variants, sparse stochastic formulations, and multilayer models for endmember variability (Kouakou et al., 13 Sep 2025, Lin et al., 2017, Abdolali et al., 2024, Tatli et al., 2023, Liu et al., 2024).

1. Canonical model and geometric interpretation

In its standard form, SSMF observes a data matrix $X \in \mathbb{R}^{m \times n}$ and seeks factors $W \in \mathbb{R}^{m \times r}$ and $H \in \mathbb{R}^{r \times n}$ such that

$X \approx W H,$

with each column $H(:,j)$ constrained to the simplex

$$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$

Equivalently, the columns of $X$ lie in $H$0, the convex hull of the columns of $H$1 (Abdolali et al., 2024). In hyperspectral unmixing, the same structure is commonly written as

$H$2

where $H$3 is an endmember or basis matrix, $H$4 is an abundance matrix, and the columns of $H$5 lie in the simplex (Liu et al., 2024, Fu et al., 2016).

The simplex constraint is the central structural assumption. It differs from generic low-rank factorization by enforcing nonnegativity and unit-sum constraints on the mixture coefficients, thereby turning the latent representation into a convex combination. This is why SSMF is often presented as a generalization of nonnegative matrix factorization, but with an explicitly geometric interpretation in which the latent factors define a simplex enclosing the data (Abdolali et al., 2024, Thanh et al., 2022).

Two classical special cases organize much of the literature. The first is the pure-pixel or separable setting, in which some data columns coincide with simplex vertices. The second is the non-separable setting, in which no such exact vertex samples are available and identification must rely on geometric regularity, minimum-volume principles, sufficiently scattered conditions, or facet structure (Lin et al., 2017, Abdolali et al., 2020).

2. Identifiability regimes and recovery conditions

A central question in SSMF is when the factorization is unique, up to permutation. Several distinct identifiability regimes appear in the literature, and they differ in both their assumptions and their algorithmic consequences.

Minimum-volume methods seek the smallest simplex that encloses the data. In the noiseless setting, a canonical formulation is

$H$6

or a closely related determinant surrogate (Kouakou et al., 13 Sep 2025, Fu et al., 2016). These methods are tied to the sufficiently scattered condition (SSC), under which identifiable solutions can be obtained without pure pixels. In the dual-volume formulation, the SSC is stated as

$H$7

along with an orthogonality exclusion condition, and under this assumption the global optimum of the dual problem recovers the facet-normals of the simplex up to column permutation (Abdolali et al., 2024).

The minimum-volume and maximum-volume-inscribed-ellipsoid lines of work share a common recovery threshold. In the MVIE framework, the uniform-pixel-purity level

$H$8

controls exact recovery, and the theorem states that if $H$9 and $W$0, then the unique MVIE inside $W$1 yields the original vertices through contact-point reconstruction (Lin et al., 2017). The VolMin literature also states a radius-based condition $W$2, and proves that this is equivalent to a cone-based sufficient condition (Fu et al., 2016).

A different line of work replaces spread conditions by sparsity on simplex facets. The facet-based condition (FBC) requires that each $W$3-facet of the true simplex contain at least $W$4 distinct columns of $W$5, while any facet of $W$6 that is not a facet of the true simplex contains strictly fewer than $W$7 data points (Abdolali et al., 2020). Under FBC, both a brute-force polytope-identification method and a greedy facet-identification method are provably correct in the noiseless setting. This condition is presented as weaker than SSC in most cases, particularly when the data are sparse on simplex facets rather than well spread throughout the interior (Abdolali et al., 2020).

Bounded SSMF introduces a different identifiability mechanism. In BSSMF, the entries of each column of $W$8 are constrained to prescribed intervals $W$9, and $X \in \mathbb{R}^{m \times n}$0 remains column-stochastic. The corresponding theorem states that if both

$X \in \mathbb{R}^{m \times n}$1

satisfy SSC, then the decomposition is essentially unique (Thanh et al., 2022). This contrasts with the statement that unbounded SSMF without volume regularization is never identifiable (Thanh et al., 2022).

Framework	Identifiability mechanism	Representative result
VolMin / MVCU	Minimum-volume enclosing simplex	Exact recovery under $X \in \mathbb{R}^{m \times n}$2 (Fu et al., 2016)
MVIE	Maximum-volume ellipsoid inscribed in $X \in \mathbb{R}^{m \times n}$3	Exact recovery under $X \in \mathbb{R}^{m \times n}$4 (Lin et al., 2017)
Dual volume maximization	Maximum-volume simplex in the polar polytope	Recovery of facet-normals under SSC (Abdolali et al., 2024)
Facet-based methods	Many data points on each true facet	Exact recovery under FBC (Abdolali et al., 2020)
BSSMF	Bounds on $X \in \mathbb{R}^{m \times n}$5 plus SSC on stacked factors	Essential uniqueness (Thanh et al., 2022)

These regimes are not interchangeable. A plausible implication is that different SSMF instances are best characterized by different geometric signatures: interior spread, facet occupancy, boundedness of latent prototypes, or availability of near-vertex samples.

3. Algorithmic families

The algorithmic literature on SSMF is correspondingly heterogeneous. Minimum-volume constrained unmixing (MVCU) remains one of the dominant paradigms. A canonical formulation minimizes a volume proxy such as $X \in \mathbb{R}^{m \times n}$6, subject to exact or approximate factorization and simplex constraints on $X \in \mathbb{R}^{m \times n}$7 (Kouakou et al., 13 Sep 2025). Classical MVCU solvers such as MVES, MVSA, and SISAL operate in batch mode and are widely used in offline settings (Kouakou et al., 13 Sep 2025).

Robust VolMin methods introduce a data-fitting term together with a log-determinant regularizer and iterative reweighting for outlier resistance: $X \in \mathbb{R}^{m \times n}$8 The updates are obtained through majorization-minimization or block successive upper-bound minimization (BSUM), with closed-form or projected-gradient substeps for the factors (Fu et al., 2016).

The Bayesian determinant-minimization perspective provides a probabilistic interpretation for this class of objectives. With Gaussian noise, simplex-uniform latent coefficients, Gaussian row priors on $X \in \mathbb{R}^{m \times n}$9, and an inverse-Wishart prior on the covariance, the MAP estimator becomes

$W \in \mathbb{R}^{m \times r}$0

with $W \in \mathbb{R}^{m \times r}$1 after concentrating out the covariance (Tatli et al., 2023). This connects determinant regularization directly to a hierarchical probabilistic model rather than treating it only as a geometric heuristic.

The MVIE framework follows a two-stage pipeline: affine dimension reduction, facet enumeration of the reduced polytope, convex log-det maximization with second-order-cone constraints, contact-point extraction, and reconstruction of the simplex vertices (Lin et al., 2017). Its main algorithmic advantage is convexity after facet enumeration. Its principal bottleneck is that facet enumeration can be worst-case exponential in the number of facets (Lin et al., 2017).

Dual simplex volume maximization transforms primal minimum-volume SSMF into a maximum-volume problem on the polar polytope. In the noisy setting it introduces slack variables and solves a determinant-based objective by BSUM, cycling over columns of a lifted matrix $W \in \mathbb{R}^{m \times r}$2 and solving small-to-medium quadratic programs (Abdolali et al., 2024). BSUM guarantees that the objective increases monotonically and converges to a stationary point (Abdolali et al., 2024).

Sparse stochastic matrix factorization imposes simplex and $W \in \mathbb{R}^{m \times r}$3-sparsity constraints on both factors and uses a column-wise update algorithm. Each update either accepts an exact block-descent point or performs a projected-gradient step onto the intersection of a simplex and a cardinality constraint, with global convergence to a critical point established via the Kurdyka-\L{}ojasiewicz property (Xiao et al., 2021).

BSSMF uses an inertial block-coordinate descent of TITAN type, alternating projections onto interval boxes for $W \in \mathbb{R}^{m \times r}$4 and simplex projections for $W \in \mathbb{R}^{m \times r}$5, and extends directly to missing-data settings by replacing the loss with $W \in \mathbb{R}^{m \times r}$6 (Thanh et al., 2022). Online SSMF, by contrast, wraps any offline MVCU solver and updates only when a new sample is geometrically informative with respect to the current simplex estimate (Kouakou et al., 13 Sep 2025).

4. Multilayer SSMF and endmember variability

The multilayer SSMF model for hyperspectral unmixing addresses endmember variability (EV), namely the fact that the endmember matrix factor may vary from one pixel to another. The starting point is the linear mixture model

$W \in \mathbb{R}^{m \times r}$7

or in matrix form,

$W \in \mathbb{R}^{m \times r}$8

with simplex constraints

$W \in \mathbb{R}^{m \times r}$9

The key modeling postulate is that if the varied endmembers are arranged as an expanded endmember matrix, then that matrix exhibits a low-rank structure (Liu et al., 2024).

The single-bundle expansion writes

$H \in \mathbb{R}^{r \times n}$0

where $H \in \mathbb{R}^{r \times n}$1, $H \in \mathbb{R}^{r \times n}$2, and the columns of $H \in \mathbb{R}^{r \times n}$3 lie in the simplex. Hence each expanded endmember column is a convex combination of “core” columns in $H \in \mathbb{R}^{r \times n}$4 (Liu et al., 2024). Substituting this into the linear mixture model yields a two-layer factorization,

$H \in \mathbb{R}^{r \times n}$5

and, more generally, an $H \in \mathbb{R}^{r \times n}$6-layer model,

$H \in \mathbb{R}^{r \times n}$7

where $H \in \mathbb{R}^{r \times n}$8 is the core basis matrix and each $H \in \mathbb{R}^{r \times n}$9 has columns in the unit simplex. Defining $X \approx W H,$0 makes $X \approx W H,$1 the expanded endmember matrix (Liu et al., 2024).

The probabilistic model assumes i.i.d. top-layer abundances

$X \approx W H,$2

and i.i.d. Gaussian noise

$X \approx W H,$3

The unknown parameters are

$X \approx W H,$4

with feasibility constraints $X \approx W H,$5, simplex constraints on each $X \approx W H,$6, and $X \approx W H,$7 (Liu et al., 2024). Exact maximum likelihood is intractable because

$X \approx W H,$8

has no closed form. The method therefore introduces variational Dirichlet distributions $X \approx W H,$9 and maximizes an ELBO obtained by Jensen’s inequality (Liu et al., 2024).

The resulting algorithm alternates two blocks. In the E-step, each $H(:,j)$0 is updated using standard Dirichlet ELBO formulas. In the M-step, $H(:,j)$1 and the $H(:,j)$2 are updated by accelerated proximal gradient descent, enforcing nonnegativity on $H(:,j)$3 and simplex constraints on the $H(:,j)$4; projection onto the simplex can be done via Condat’s method. The noise variance $H(:,j)$5 has a closed-form update from the average residual variance. Initialization proceeds from one-layer HU, such as VCA+FCLS, followed by random Dirichlet initialization for the intermediate simplex factors (Liu et al., 2024).

The empirical evaluation in the paper is organized around both synthetic and real data. On the Samson image, a $H(:,j)$6 sub-image with $H(:,j)$7 bands and $H(:,j)$8 materials, the multilayer model uses dimensions $H(:,j)$9 and is benchmarked against ELMM and AEB. The reported abundance maps show finer spatial detail, and the hierarchical endmember layers recover three “core” signatures plus three perturbation bases; subsequent layers generate $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$0 expanded signatures through convex combinations, capturing EV in a material-specific manner, with water most variable and tree least (Liu et al., 2024). On synthetic data with $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$1 bands, $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$2 base endmembers, $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$3 library variants, $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$4, and model dimensions $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$5, the metric

$$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$6

decreases smoothly with SNR, and the extracted spectra closely match the ground truth (Liu et al., 2024).

Within SSMF more broadly, this multilayer construction is notable because it models pixel-wise EV without an external library. The hierarchy of convex-combination factors suggests a mechanism for separating base signatures from perturbation modes while preserving the simplex structure at every layer.

5. Variants, large-scale settings, and application domains

SSMF has expanded beyond the classical batch minimum-volume setting. One direction is online processing. The online SSMF framework oSSMF is designed for sequential data and wraps an offline MVCU routine, invoking it only when a new sample lies outside the current estimated simplex or near one of its facets. It stores only observations deemed informative with respect to the geometric constraints of SSMF (Kouakou et al., 13 Sep 2025). On synthetic data with $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$7, $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$8, $$\Delta^r = \{ h \in \mathbb{R}^r : h \ge 0,\ \sum_{i=1}^r h_i = 1 \}.$$9, and on a Raman hyperspectral image of size $X$0 with $X$1, oSSMF nearly matches offline SISAL in aSAD and RMSE while reducing runtime per iteration from $X$2 s to $X$3 s on synthetic data and from $X$4 s to $X$5 s on real data, corresponding to a $X$6 and $X$7 speed-up, respectively (Kouakou et al., 13 Sep 2025).

Another direction is bounded latent geometry. BSSMF constrains the entries of $X$8 to lie in given intervals, so that $X$9 inherits the same dynamic range. This is especially natural when entries of $H$00 themselves belong to prescribed intervals, such as pixel intensities in $H$01 or ratings in $H$02 (Thanh et al., 2022). On MNIST, BSSMF$H$03 produces basis vectors that look like whole digit prototypes, whereas NMF basis vectors look like parts of digits. On recommender-system matrix completion, BSSMF remains stable as the rank grows while unconstrained MF and NMF overfit more strongly; on MovieLens-1M at rank $H$04, test-RMSE is reported as $H$05 for BSSMF, $H$06 for NMF, and $H$07 for unconstrained MF (Thanh et al., 2022).

Facet-based SSMF is particularly effective in highly mixed, noisy, rank-deficient, or outlier-contaminated settings. On synthetic data, GFPI achieves zero reconstruction error up to machine precision in noiseless experiments across a range of purity levels, while SNPA fails as soon as purity drops below $H$08 and min-vol or MVIE degrade at low purity. In noisy experiments, GFPI yields the lowest relative error across purity values (Abdolali et al., 2020). On hyperspectral benchmarks, Samson and Jasper Ridge, GFPI attains MRSA values of $H$09 and $H$10, respectively; on Jasper Ridge it outperforms SNPA, Min-Vol, and HyperCSI in MRSA (Abdolali et al., 2020).

VolMin methods have also been applied outside remote sensing. In document clustering on Reuters21578, a robust VolMin formulation with $H$11 reaches clustering accuracy $H$12 for $H$13, compared with $H$14 for VolMin with $H$15 and $H$16 for LCCF (Fu et al., 2016). The same paper emphasizes remote sensing and document clustering as two application domains for volume-minimization-based structured matrix factorization (Fu et al., 2016). Topic modeling is treated explicitly in the dual-volume work, which states that the same dual formulation can be applied to word-document count matrices after appropriate preprocessing (Abdolali et al., 2024).

6. Limitations, controversies, and open problems

Despite the breadth of formulations, several limitations recur across the literature. Minimum-volume models are non-convex in the primal factors and typically guarantee only local solutions unless additional structure is exploited (Fu et al., 2016). Robust VolMin requires tuning of $H$17, $H$18, $H$19, and $H$20 (Fu et al., 2016). MVIE avoids spurious local minima because the core optimization problem is convex, but facet enumeration can be expensive and worst-case exponential (Lin et al., 2017). Dual simplex volume maximization has strong identifiability guarantees under SSC, yet its noisy optimization procedure remains non-convex and motivates open questions about the global landscape and specialized solvers (Abdolali et al., 2024).

Online SSMF scales well, but its current formulation explicitly lacks formal convergence proofs and worst-case guarantees. It also requires tuning of tolerances $H$21, a facet threshold $H$22, and a proximity threshold $H$23, and informative samples may be missed in extreme noise or highly non-convex sampling (Kouakou et al., 13 Sep 2025). Sparse stochastic formulations provide global convergence to a critical point, but the same source notes that factorization is not unique in general without extra separability or anchor-word conditions, and automatic choice of sparsity levels remains challenging (Xiao et al., 2021).

There is also a methodological tension between spread-based and facet-based views of identifiability. SSC requires the data to be well spread inside the simplex, whereas facet-based conditions exploit concentration of data on lower-dimensional faces (Abdolali et al., 2020, Abdolali et al., 2024). These are not contradictory, but they privilege different empirical regimes. This suggests that no single identifiability condition dominates across all applications.

For hyperspectral unmixing, endmember variability creates an additional difficulty because the effective mixing geometry is no longer captured by a single fixed endmember matrix. The multilayer SSMF model addresses this by imposing a low-rank structure on an expanded endmember matrix and factorizing it through multiple simplex-structured layers (Liu et al., 2024). A plausible implication is that future SSMF research will continue to move from static simplex recovery toward hierarchical, streaming, bounded, and uncertainty-aware models, while retaining the core geometric principle that mixture coefficients lie in a simplex.