Identifiability of the Simplex Volume Minimization Criterion for Blind Hyperspectral Unmixing: The No Pure-Pixel Case

1 Introduction

• MVES minimizes the volume of a simplex enclosing the hyperspectral data, using its vertices to estimate endmember signatures.The criterion is nonconvex, and practical algorithms had been developed, but this paper focuses on identifiability rather than algorithm design.Source: §1 · p. 2
• The study addresses the gap between empirical reports of accurate MVES recovery without pure pixels and the absence of a theoretical guarantee specifying how much mixing it can withstand.The paper asks where the identifiability limit lies as mixing increases.Source: §1 · p. 2
• The analysis covers the no-pure-pixel case for any number of endmembers and proves exact, unique recovery under a condition on pixel purity and geometry.For three or more endmembers, the condition is described as much milder than the pure-pixel assumption; for two, it is equivalent to that assumption.Source: §1 · pp. 2–3

2.1 Preliminaries

• A simplex is an (N −1)-dimensional convex hull of N affinely independent vertices; it is regular when all vertex-to-vertex distances are equal.The paper also defines its volume using the simplex vertices.Source: §2.1 · p. 3

2.2 Blind HU Problem Setup

• In the noiseless setup, each observed pixel is a linear mixture of endmember signatures, with nonnegative abundances summing to one.The formulation assumes full column rank for the endmember matrix, full row rank for the abundance matrix, and known endmember count.Source: §2.2 · pp. 3–4

2.3 Minimum-Volume Enclosing Simplex

• MVES estimates the endmember matrix by finding a minimum-volume simplex that encloses all observed pixels.The paper treats the true endmember matrix as the target when the optimal solution is exact and unique, up to permutation.Source: §2.3 · p. 4
• The central question is whether MVES uniquely identifies the true endmembers when no abundance vector is a pure pixel.Pure-pixel identifiability was already known; empirical evidence suggested recovery could also occur without pure pixels.Source: §2.3 · p. 4

3.1 Pixel Purity Measures

• The best purity level ρ reflects the purest abundance vector, while the uniform level γ captures purity across all endmembers and distinguishes symmetric from asymmetric spreads.The two levels can be similar for relatively symmetric abundance distributions and diverge when purity is uneven across endmembers.Source: §3.1 · pp. 5–6

3.2 Provable MVES Identifiability

• Exact MVES identifiability for the observed pixels is equivalent to identifiability for their abundance vectors, and does not depend on A when A has full column rank.The paper also gives a necessary purity condition for N ≥3 and proves that γ > 1/(N−1) is sufficient for exact, unique recovery.Source: §3.2 · p. 8
• For N = 2, exact and unique MVES recovery holds if and only if the pure-pixel assumption holds.For N ≥3, the sufficient γ threshold is milder than the pure-pixel condition γ = 1.Source: §3.2 · pp. 8–9

3.3 Further Discussion

• Explicitly characterizing γ from the abundance set is difficult, and computing γ for a given set is generally NP-hard.The authors note that large samples from continuous distributions covering R(r) for r > 1/(N−1) may nevertheless yield perfect identifiability with high probability.Source: §3.3 · p. 9
• Under Assumption 1, the uniform purity level γ is characterized by the smallest αij, linking abundance-pair conditions to the sufficient identifiability criterion.For N = 3, Corollary 1 gives αij > 2/3 for all i,j; for N ≥4 it states exact identifiability under its stated condition.Source: §3.3 · p. 10
• The stated N ≥4 conclusion is conditional: satisfying Assumption 1 for general αij may require at least N(N−1) pixels.The analysis assumes noiseless data, and explicitly leaves noisy-case sensitivity for future work.Source: §3.3 · p. 10

4.1 Proof of Proposition 1

• The Proposition 1 proof uses affine-hull correspondence and a contradiction argument to relate uniqueness of the abundance-space and data-space MVESs.It invokes the full-row-rank abundance assumption when applying Lemma 1.Source: §4.1 · p. 11

4.2 Proof of Theorem 1

• For N≥3, when ρ≤1/(N−1), the proof yields infinitely many MVESs, contradicting uniqueness.The construction produces infinitely many MVESs of C(ρ), which the proof connects to those of the abundance set.Source: §4.2 · p. 14

4.3 Proof of Theorem 2

• The proof combines an enclosure relation with uniqueness for R(γ) to conclude MVES(SL)={Te}.The conclusion follows by combining the results of Steps 2 and 3.Source: §4.3 · p. 15

4.4 Proof of Theorem 3

• For 1/(N−1)<r≤1, every MVES of R(r) equals the unit simplex Te.The proof identifies an arbitrary candidate with Te using Lemma 4.Source: §4.4 · pp. 15–16

4.5 Proof of Proposition 2

• For N=2, perfect identifiability holds if and only if the pure-pixel assumption holds.In the proof’s coefficient notation, this requires pixels with αn=1 and αn=0.Source: §4.5 · p. 17

4.6 Proof of Theorem 4

• The set convP is defined as {s ∈ Te | si ≤ α, i = 1, . . . , N}, specifying vectors s in Te whose coordinates si do not exceed α.Source: §4.6 · p. 17

5 Numerical Experiments

• The experiment uses 224 spectral bands, 1,000 pixels, USGS endmembers, and 50 random realizations.Abundances are drawn from a Dirichlet distribution and filtered to control pixel purity.Source: §5 · p. 18
• Zero RMS angle error occurs when r>1/(N−1), while nonzero errors occur for r≤1/(N−1).This matches the sufficient and necessary MVES identifiability results, respectively.Source: §5 · p. 18, Figure 5 · p. 19
• The reported robustness limit is 1/(N−1) in terms of the uniform pixel purity level.The paper points to prior experimental studies for broader synthetic and real-data coverage.Source: §5 · pp. 18, 20

6 Conclusion

• The paper performs a theoretical analysis of MVES identifiability in blind hyperspectral unmixing.The analysis concerns recovery of endmembers without requiring the pure-pixel assumption.Source: §6 · p. 20
• Under some mild assumptions, MVES is robust to the lack of pure pixels.The assumptions are described as considerably more relaxed than those for the pure-pixel case.Source: §6 · p. 20
• The analysis provides a theoretical explanation for numerical studies that found accurate MVES recovery without pure pixels.Source: §6 · p. 20

A Proof of Lemma 1

• Under full column rank, the image of a simplex in the relevant affine hull is represented as the convex hull of its mapped vertices.The proof writes f(TG)=conv{h1,…,hN}.Source: A Proof of Lemma 1 · p. 20
• The simplex-volume calculation relates the determinant for mapped vertices to the original determinant and the coordinate-change determinant.The stated relation is det(ĀᵀĀ)=det(ḢᵀḢ)·|det(Θ̄)|².Source: A Proof of Lemma 1 · pp. 20–21
• For a simplex in the image affine hull, its inverse image is represented by simplex coefficients θ≥0 with 1ᵀθ=1.Full column rank of A supports the inverse-image representation.Source: A Proof of Lemma 1 · p. 21

B Proof of Lemma 3

• An enclosing simplex W is an MVES of B(µ) exactly when its corresponding V is an MVES of C(r), and their volumes are equal.The proof uses this equivalence to transfer the minimum-volume property between the two sets.Source: B Proof of Lemma 3 · p. 21
• For r = 1/(N−1), Te = conv{e1, …, eN} encloses C(r) and attains the same volume as V, making Te an MVES of C(r).The volume comparison is obtained using the simplex volume formula.Source: B Proof of Lemma 3 · p. 22

C Proof of Lemma 4

• Any W ∈ MVES(B(r)) has exactly N boundary intersections with B(r), and those points are affinely independent.These contact properties are the key geometric result used in the proof.Source: C Proof of Lemma 4 · p. 22
• The simplex W can be reconstructed from its N boundary-contact points with B(r).The proof uses this reconstruction to identify simplices once their contact sets agree.Source: C Proof of Lemma 4 · p. 22
• The proof shows the candidate contact set equals bdB(r) ∩ bdW, then uses affine independence and the reconstruction property to conclude W = W′.The argument establishes equality of the contact sets before concluding equality of the simplices.Source: C Proof of Lemma 4 · pp. 22–23

D Proof of Lemma 7

• A simplex W can be represented as P(g, H) by its vertices, and when H has full rank the representation determines its vertices and volume.This representation enables the proof to perturb simplex constraints and compare volumes.Source: D Proof of Lemma 7 · p. 24
• The ball B(r) is contained in P(g, H) when each defining inequality satisfies the stated radius-and-offset condition.The converse also holds, providing a containment test for the simplex representation.Source: D Proof of Lemma 7 · p. 24
• If B(r) ⊆ W, its boundary intersects bdW at no more than N points, each among the points specified by tight facet inequalities.The proof identifies these possible contacts from the simplex's inequality representation.Source: D Proof of Lemma 7 · pp. 24–25
• If a candidate contact is absent, the proof perturbs the simplex constraints to form a smaller-volume simplex still enclosing B(r), contradicting minimality.The same contradiction argument is applied to the final candidate contact as well.Source: D Proof of Lemma 7 · pp. 25–26
• Thus an MVES of B(r) has exactly N boundary contacts; they are affinely independent and determine W.The proof concludes the contact count by excluding missing contacts and derives the two stated properties.Source: D Proof of Lemma 7 · pp. 25–26

E Proof of Lemma 5

• For 0.5 < α < 1, the proof targets convP = U(α), with U(α) restricting simplex points to coordinates si ≤ α.The α = 1 case is immediate because both sets equal Te.Source: E Proof of Lemma 5 · p. 26
• Every s ∈ convP lies in U(α), since its coordinates are bounded by α through the coordinate bounds on the defining points pij.The proof writes s as a convex combination and applies the bounds on pij.Source: E Proof of Lemma 5 · pp. 26–27
• For the reverse inclusion, the proof orders s's coordinates, chooses k by the threshold δi, and constructs points whose convex hull contains s.It verifies the constructed points belong to convP and represents s as their convex combination.Source: E Proof of Lemma 5 · pp. 27–29
• The two inclusions establish s ∈ convP if and only if s ∈ U(α), completing the equality.The proof's coordinate and threshold properties support the convex-combination construction.Source: E Proof of Lemma 5 · p. 29

• The proof recalls R(r) as the subset of T_e whose elements have norm at most r, and begins rewriting T_e.Source: F Proof of Lemma 6 · p. 29
• For s in R(r), the proof assumes its components are ordered decreasingly and states that s_1 is at least 1/N.Source: F Proof of Lemma 6 · p. 29
• The proof reexpresses the constraint in (93b) using the quadratic formula.Source: F Proof of Lemma 6 · p. 30