Simplex-like Minimal Spectrahedra

• Simplex-like minimal spectrahedra are free convex sets defined by irreducible Arveson extreme points that serve as minimal generators, analogous to simplex vertices.
• They combine a boundary-theoretic notion of minimal generators with minimal linear matrix inequality sizes to achieve optimal convex representations.
• The methodology employs constructive dilation theory and semidefinite programming algorithms to iteratively recover absolute extreme points from free spectrahedra.

Simplex-like minimal spectrahedra arise in free convexity as compact free spectrahedra whose canonical generating boundary behaves like the vertex set of a simplex: the relevant “vertices” are the absolute extreme points, equivalently irreducible Arveson extreme points, and the whole free spectrahedron is recovered as their matrix convex hull, with minimality under inclusion among unitary-invariant irreducible generating sets (Evert et al., 2018). In parallel, classical spectrahedral theory uses “minimal” to denote the smallest matrix size in a linear matrix inequality description, so the subject combines a boundary-theoretic notion of minimal generators with an LMI-size notion of minimal representation (Kummer, 2015).

1. Matrix-convex and spectrahedral setting

Let $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$. For positive integers $g,n$, the space

$$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$

consists of $g$-tuples of self-adjoint matrices, and $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$. Given a $g$-tuple $A=(A_1,\dots,A_g)$ of self-adjoint $d\times d$ matrices, the associated monic linear pencil is

$L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$

and its evaluation at $X\in SM_n(\mathbb{F})$ is

$g,n$0

The free spectrahedron at level $g,n$1 is

$g,n$2

and the free spectrahedron is $g,n$3 (Evert et al., 2018).

The matrix-convex structure is defined by dimension-free mixing. A matrix convex combination of tuples $g,n$4 is an expression

$g,n$5

where the sizes $g,n$6 need not agree. Equivalently, if $g,n$7 and $g,n$8, then $g,n$9 with $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$0 an isometry. Thus matrix convex combinations are compressions of direct sums via isometries, and matrix convexity is the dimension-free analog of classical convexity (Evert et al., 2018).

This setting is the natural ambient space for “simplex-like” questions. Classical simplices are characterized by finitely many vertices and scalar convex combinations, whereas free spectrahedra are organized across all matrix sizes and use partial isometries rather than scalar weights. A plausible implication is that any simplex analogy in this setting must be expressed through boundary points stable under compression, rather than through ordinary extreme points at a single level.

2. Extremal boundary points and the noncommutative vertex picture

For matrix convex sets there are several extremal notions. In the spectrahedral setting the decisive ones are Arveson extreme points and absolute extreme points. A tuple $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$1 is an Arveson extreme point if every dilation

$$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$2

has $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$3. Equivalently, if $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$4 with $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$5 and $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$6 an isometry, then $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$7. An absolute extreme point is one that admits no nontrivial weakly proper matrix convex combination except by unitary copies of itself, possibly with extra direct summands (Evert et al., 2018).

Notion	Characterization	Role
Arveson extreme point	No nontrivial dilation inside $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$8	Boundary notion tied to dilations
Absolute extreme point	No nontrivial weakly proper matrix convex decomposition except unitary copies/direct-sum extensions	Minimal generating “vertex”
Irreducible tuple	No nontrivial common reducing subspace for $$SM_n(\mathbb{F})=\{X=(X_1,\dots,X_g): X_j^*=X_j\in M_n(\mathbb{F})\}$$9	Irreducibility condition for absolute extremality

A key structural identification is: $g$0 For free spectrahedra closed under complex conjugation, this theorem identifies the correct noncommutative analog of a vertex: Arveson extremality captures boundary rigidity under dilations, while irreducibility removes reducible direct-sum artifacts (Evert et al., 2018).

The simplex analogy is precise but not literal. In a classical simplex, vertices are ordinary extreme points and support unique barycentric coordinates. In a free spectrahedron, the analog of a vertex is an irreducible Arveson extreme point, but matrix convex representations need not be unique. The analogy lies in minimal generation, not in uniqueness of coordinates.

3. Spanning, minimality, and constructive dilation theory

The central theorem states that every compact free spectrahedron is spanned by its absolute extreme points: $g$1 More strongly, if $g$2 is any closed matrix convex set and $g$3 is a set of irreducible tuples that is closed under unitary conjugation and satisfies $g$4, then

$g$5

Thus the absolute boundary is the smallest unitary-invariant irreducible generating set whose matrix convex hull is $g$6 (Evert et al., 2018).

This is the precise sense in which compact free spectrahedra are simplex-like. The paper explicitly interprets the result as the direct free analog of the statement that the vertices of a simplex form the unique minimal generating set of its convex hull. The analogy is strengthened by finite representation bounds. If $g$7 and $g$8, where

$g$9

then there exists an Arveson extreme dilation of $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$0 with controlled extra size. In the real case one has

$SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$1

while in the complex case

$SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$2

Consequently, every $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$3 is a matrix convex combination of absolute extreme points with total size bounded by $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$4 (Evert et al., 2018).

The construction is algorithmic in the real case. The dimension of the dilation subspace measures failure of Arveson extremality: $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$5 When $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$6 is not Arveson extreme, one performs maximal $SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$7-dilations

$SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$8

and Theorem 2.3 gives the strict dimension drop

$SM(\mathbb{F})=\bigcup_{n\ge1}SM_n(\mathbb{F})$9

Iterating yields an Arveson extreme $g$0-dilation in at most $g$1 steps. The paper formulates each step as an SDP followed by a local maximization over the lower-right block, so the spanning theorem is constructive rather than merely existential (Evert et al., 2018).

4. Classical analogies, Choquet-type structure, and failure beyond spectrahedra

The paper places these results alongside Krein–Milman and Carathéodory. Classically, a compact convex set is the closed convex hull of its extreme points, and points in $g$2 admit convex combinations of at most $g$3 extreme points. In the free setting, the analogue is a finite matrix-convex representation by absolute extreme points, with explicit total-size bounds rather than a fixed ambient-dimension bound. The paper describes this as a free Choquet–Carathéodory statement (Evert et al., 2018).

The analogy has sharp limits. Representations by absolute extreme points need not be unique; the geometry is multilevel rather than controlled by a single affine dimension; and matrix convex combinations use compressions of direct sums rather than scalar weights. Moreover, for general matrix convex sets not defined by finite-dimensional LMIs, the analog can fail. The spectrahedral hypothesis is therefore structural, not cosmetic (Evert et al., 2018).

Later work sharpened these limitations. Over the complexes, bounded free spectrahedra closed under complex conjugation satisfy a spanning theorem by free extreme points, but this does not extend to free spectrahedrops: there are closed, bounded free spectrahedrops that are closed under complex conjugation and have no free extreme points. Over the reals, the free polar dual of a bounded real free spectrahedron is rarely the projection of a real free spectrahedron, and if such a projection exists under the paper’s conjugation-closure hypothesis, then the first level must be a polyhedron (Evert et al., 27 Jul 2025).

These results show that simplex-like minimality is most robust at the level of free spectrahedra themselves. A plausible implication is that projection, even when it preserves ordinary convexity at level $g$4, can destroy the boundary structure that makes a spectrahedron behave like a noncommutative simplex.

5. Polyhedrality, diagonal normal forms, and minimal LMI size

A second notion of minimality concerns the size of spectrahedral descriptions. For a classical spectrahedron

$g$5

one asks for the smallest matrix size $g$6 among all such LMI representations. For compact convex sets defined by a quadratic polynomial, if $g$7 then $g$8; in particular $g$9, and if $A=(A_1,\dots,A_g)$0 then $A=(A_1,\dots,A_g)$1. For convex regions in $A=(A_1,\dots,A_g)$2 with smooth cubic algebraic boundary, no size-$A=(A_1,\dots,A_g)$3 description exists, so the minimal size is at least $A=(A_1,\dots,A_g)$4 (Kummer, 2015).

Polyhedrality provides a complementary structural criterion. A full-dimensional spectrahedral cone $A=(A_1,\dots,A_g)$5 is polyhedral if and only if, after a congruence, one has a block decomposition

$A=(A_1,\dots,A_g)$6

with $A=(A_1,\dots,A_g)$7 diagonal and

$A=(A_1,\dots,A_g)$8

The diagonal block contains all facet-defining linear inequalities, while the proper block $A=(A_1,\dots,A_g)$9 is the strictly non-polyhedral part (Bhardwaj et al., 2011). This suggests that, in the classical polyhedral regime, a simplex-like spectrahedron is reflected by a purely diagonal normal form.

The same polyhedral viewpoint appears in spectral convexity. If $d\times d$0 is a symmetric polyhedron, the spectral polyhedron

$d\times d$1

is a spectrahedron. If $d\times d$2 has $d\times d$3 orbits of facets, the explicit representation has size

$d\times d$4

while any spectrahedral representation of $d\times d$5 must have size at least $d\times d$6. If the origin lies in the interior of $d\times d$7, then $d\times d$8 is the projection of a spectrahedron of size

$d\times d$9

These bounds show that “minimal spectrahedron” may also mean minimal defining pencil or minimal lifted size, not only minimal generating boundary (Sanyal et al., 2020).

6. Euclidean balls, failed global models, and local simplicial faces

The Euclidean ball is a decisive test case for simplex-like behavior. Let $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$0 denote the universal anticommuting self-adjoint unitaries. For $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$1,

$L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$2

so the disk has a particularly clean minimal/maximal free convex model. For $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$3, however,

$L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$4

because $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$5 has higher-level free extreme points, and these propagate to all $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$6. The paper concludes that these phenomena obstruct a simplex-like minimal spectrahedron for the Euclidean ball in dimension $L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$7 and higher (Evert et al., 27 Jul 2025).

The failure is not merely representational. In the real setting, if the free polar dual of a bounded real free spectrahedron is the projection of a real free spectrahedron under the stated conjugation-closure hypothesis, then the first level must be a polyhedron. Since the Euclidean ball is not a polyhedron, its real free polar dual cannot arise in that way (Evert et al., 27 Jul 2025). This sharply separates polyhedral, simplex-related behavior from ball-like free convexity.

A different development shows that simplex-like structure can nevertheless appear locally. In the homogeneous pseudo-moment cone

$L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$8

if

$L_A(x)=I_d+\sum_{j=1}^g A_jx_j,$9

with

$X\in SM_n(\mathbb{F})$0

then for generically chosen atoms the minimal face satisfies

$X\in SM_n(\mathbb{F})$1

and these generators are linearly independent. The resulting minimal face is simplicial and generated by the planted rank-one atoms, yielding a Carathéodory-type atomic decomposition algorithm that exactly recovers them in the generic regime (Kang et al., 7 May 2026).

This suggests a broader interpretation of simplex-like minimal spectrahedra. Globally, free spectrahedra may fail to resemble simplices, especially under projection or for curved first levels such as Euclidean balls. Locally, however, minimal faces of structured spectrahedral cones can become simplicial, so the simplex analogy survives as a statement about the smallest face containing a point rather than about the entire convex body.