Fusion-Space Codes: Optimal Subspace Packings

• Fusion-space codes are collections of subspaces in a finite-dimensional Hilbert space selected to maximize separation under prescribed metrics.
• They employ structures like tight fusion frames, equi-isoclinic tight fusion frames, and mixed-rank embeddings, constructed via harmonic analysis and Radon-Hurwitz theory.
• These codes are pivotal in applications such as compressed sensing, MIMO communications, and quantum information, with explicit constructions using group theory and design methods.

Searching arXiv for the cited papers and closely related work on fusion-space codes. arXiv search query: "Fusion-space codes Grassmannian equi-isoclinic tight fusion frame harmonic mixed-rank Radon-Hurwitz" Fusion-space codes are Grassmannian codes whose codewords are subspaces, or equivalently orthogonal projections onto subspaces, of a finite-dimensional Hilbert space. In the constant-rank setting, they are studied as packings of $N$ $R$-dimensional subspaces in $\mathbb{R}^D$ or $\mathbb{C}^D$ that maximize separation under a prescribed subspace metric; in the mixed-rank setting, they generalize this problem to collections of different dimensions. The modern theory is organized around tight fusion frames, equi-isoclinic and equichordal structure, embedding methods for comparing subspaces of unequal rank, and explicit constructions from harmonic analysis, finite abelian groups, mutually unbiased bases, block designs, and Radon-Hurwitz theory (Fickus et al., 2021, Casazza et al., 2019, Fickus et al., 2024).

1. Basic objects and terminology

A fusion-space code, also called a Grassmannian code in the cited literature, is a collection of subspaces of a fixed ambient Hilbert space chosen to maximize the minimal distance between distinct codewords. In the constant-rank case, the code consists of $N$ subspaces $\{U_n\}_{n=1}^N$, each of dimension $R$, inside $\mathbb{C}^D$ or $\mathbb{R}^D$. Each subspace is represented by its orthogonal projection $P_n$, and a central structural condition is the tight fusion frame condition

$R$0

A sequence satisfying this identity is a tight fusion frame (TFF) (Fickus et al., 2021).

Two refinements of TFFs are fundamental. An equichordal tight fusion frame (ECTFF) is a TFF for which the sum of squares of cosines of principal angles is constant among all distinct pairs. An equi-isoclinic tight fusion frame (EITFF) is a TFF in which every pair of distinct subspaces has all principal angles equal. In projection language, equi-isoclinicity means that for all $R$1,

$R$2

for a constant $R$3 independent of the pair. EITFFs are the central optimal objects for spectral-distance packing and for minimizing block coherence in block compressed sensing (Fickus et al., 2021).

The mixed-rank theory reformulates the same problem when the subspaces are allowed to have different dimensions. In that setting, a packing is parameterized by pairs $R$4, meaning $R$5 subspaces of rank $R$6, with $R$7. This resolves the comparison problem for unequal dimensions by embedding all projections into a common Euclidean sphere and defining optimality there (Casazza et al., 2019).

2. Distances, coherence, and optimality criteria

The constant-rank theory uses several equivalent measures of separation. For two $R$8-dimensional subspaces $R$9 and $\mathbb{R}^D$0 with projections $\mathbb{R}^D$1 and $\mathbb{R}^D$2, the chordal distance is

$\mathbb{R}^D$3

while the spectral distance is

$\mathbb{R}^D$4

where $\mathbb{R}^D$5 are the principal angles. The spectral distance is controlled by the smallest principal angle, whereas the chordal distance aggregates all principal angles (Fickus et al., 2021).

For block-sparse signal models, the relevant quantity is block coherence,

$\mathbb{R}^D$6

EITFFs minimize this value. The governing bound is the generalized Welch bound: $\mathbb{R}^D$7 Equality holds if and only if the fusion frame is equi-isoclinic and tight. Accordingly, an EITFF is optimal as a Grassmannian code with respect to spectral distance (Fickus et al., 2021).

The mixed-rank theory replaces direct comparison of subspaces by a traceless embedding. For a rank-$\mathbb{R}^D$8 projection $\mathbb{R}^D$9 on $\mathbb{C}^D$0, the embedding is

$\mathbb{C}^D$1

After normalization, all embedded images lie on a common sphere, and optimality is defined by minimizing the maximal pairwise inner product among the embedded points. The resulting problem is a restricted coding problem on a union of compact subsets of the sphere. In this sense, mixed-rank fusion-space coding extends the coherence-minimization program of constant-rank packings to heterogeneous subspace families (Casazza et al., 2019).

3. Harmonic constructions and operator-valued difference sets

A major construction paradigm is harmonic. A harmonic TFF is a TFF whose Gram matrix is block-circulant under the action of a finite abelian group $\mathbb{C}^D$2. Equivalently, the arrangement can be realized as the orbit of a single subspace under a unitary representation of $\mathbb{C}^D$3. This generalizes the classical harmonic ETF/difference-set correspondence from rank $\mathbb{C}^D$4 to higher-rank fusion frames (Fickus et al., 2021).

The harmonic framework is formulated in terms of a family of projections $\mathbb{C}^D$5 on $\mathbb{C}^D$6, possibly of varying ranks $\mathbb{C}^D$7. Its matrix-valued discrete Fourier transform is

$\mathbb{C}^D$8

The associated harmonic TFF is an EITFF if and only if there exists $\mathbb{C}^D$9 such that

$N$0

Equivalently, the projections satisfy

$N$1

for some scalar $N$2. This operator-valued condition is the fusion-frame analogue of a difference set; the paper terms such data operator-valued difference sets, or difference projections. For $N$3, the condition reduces to the classical difference-set condition (Fickus et al., 2021).

The explicit constructions exploit Gauss sums over finite fields,

$N$4

with additive character $N$5 and multiplicative character $N$6. Their magnitude and symmetry properties guarantee that the Fourier transforms $N$7 are scalar multiples of unitaries, yielding equi-isoclinic harmonic packings. The paper constructs EITFFs consisting of $N$8 planes in $N$9 for each prime power $\{U_n\}_{n=1}^N$0, of $\{U_n\}_{n=1}^N$1 planes in $\{U_n\}_{n=1}^N$2 for each odd prime power $\{U_n\}_{n=1}^N$3, and of $\{U_n\}_{n=1}^N$4 three-dimensional subspaces in $\{U_n\}_{n=1}^N$5 (Fickus et al., 2021).

These constructions are notable because most previously known EITFFs had parameters matching direct-sum or “tensor-sized” constructions derived from equiangular tight frames. The harmonic method produces new infinite families of non-“tensor-sized” EITFFs and supplies a unified Fourier-analytic framework for their study (Fickus et al., 2021).

4. Radon-Hurwitz codes in the half-dimensional case

A second major structural result concerns the special regime $\{U_n\}_{n=1}^N$6, where the subspace dimension is exactly one-half of the ambient dimension. In this case, EITFFs are fully characterized by Radon-Hurwitz theory. The paper shows that every EITFF$\{U_n\}_{n=1}^N$7 is equivalent to one built from isometries

$\{U_n\}_{n=1}^N$8

with

$\{U_n\}_{n=1}^N$9

where the $R$0 are unitary $R$1 matrices satisfying

$R$2

Conversely, any such family of unitaries yields an EITFF$R$3 (Fickus et al., 2024).

Existence is controlled exactly by the Radon-Hurwitz number. If $R$4 denotes the maximal size of a $R$5-orthonormal family in $R$6 matrices over $R$7, then an EITFF$R$8 exists if and only if

$R$9

This sharp characterization refines earlier bounds and identifies the half-dimensional case as one where optimal fusion-space codes are governed by classical algebraic-topological structure (Fickus et al., 2024).

The same paper proves a strong symmetry phenomenon. Every such “Radon-Hurwitz EITFF” is highly symmetric, and every even permutation is an automorphism. In many cases the symmetry group is larger; the paper distinguishes between alternating symmetry and total symmetry and shows that the latter is tied to additional algebraic conditions on the underlying $\mathbb{C}^D$0. A central implication is that in this regime optimality forces substantial symmetry rather than merely coexisting with it (Fickus et al., 2024).

5. Mixed-rank packings and combinatorial constructions

The mixed-rank theory addresses a longstanding open problem by reformulating Grassmannian fusion frames to the case of mixed dimensions and showing that the resulting notion has the proper properties for the problem. The traceless embedding gives all subspaces “equal footing,” after which the packing problem becomes one of placing points optimally on a higher-dimensional sphere subject to membership in a union of compact sets (Casazza et al., 2019).

Within this framework, an embedded mixed-rank packing is optimally spread if its maximum pairwise inner product is minimal among all packings with the prescribed ranks and multiplicities. Several structural consequences follow. Every optimally spread mixed-rank packing is a fusion frame. Optimality is preserved under spatial complementation $\mathbb{C}^D$1. When the embedded vectors form an orthoplex and the number of subspaces reaches the maximal value $\mathbb{C}^D$2, the resulting objects are maximal orthoplectic fusion frames, and such maximal orthoplectic fusion frames are always tight (Casazza et al., 2019).

The principal explicit constructions use mutually unbiased bases (MUBs) and block designs. For a basis $\mathbb{C}^D$3 and an index subset $\mathbb{C}^D$4, the coordinate projection is

$\mathbb{C}^D$5

Using maximal sets of MUBs together with block designs having controlled intersection sizes, one obtains infinite families of tight, optimally spread mixed-rank fusion frames. Affine designs yield families with two different ranks. Symmetric block designs, including projective planes, Menon designs, and Hadamard designs, produce further examples. A particularly strong characterization states that maximal orthoplectic constant-rank fusion frames arise if and only if the construction comes from Hadamard 3-designs (Casazza et al., 2019).

This mixed-rank program changes the scope of fusion-space coding. The constant-rank setting remains central, but the general theory shows that optimal packings need not be restricted to a single Grassmannian manifold; instead, they may live in a product of Grassmannians while still retaining a precise spherical-coding interpretation (Casazza et al., 2019).

6. Applications, scope, and conceptual distinctions

Fusion-space codes are studied because optimal subspace separation is useful in compressed sensing, signal processing, coding theory, quantum information, MIMO communications, radar, and multiuser communication. In the constant-rank case, EITFFs yield dictionaries with minimal block coherence and therefore optimal guarantees for block-sparse recovery under the cited model. In the mixed-rank case, the embedding formulation supplies a geometric language for resilient and heterogeneous subspace arrangements, with explicit structured constructions from MUBs and block designs (Fickus et al., 2021, Casazza et al., 2019).

Several distinctions are essential. First, equi-isoclinic and equichordal are not interchangeable: EITFFs require equality of all principal angles for each pair, whereas ECTFFs only require constancy of a summed quantity derived from those angles (Fickus et al., 2021). Second, optimality depends on the metric and the rank model. In constant rank, spectral-distance optimality is characterized by EITFF structure and the fusion Welch bound. In mixed rank, optimality is formulated after embedding and becomes a restricted spherical coding problem rather than a direct metric comparison inside a single Grassmannian (Casazza et al., 2019). Third, not all explicit optimal families are harmonic; the harmonic Fourier-analytic constructions and the combinatorial MUB/block-design constructions are complementary rather than identical paradigms (Fickus et al., 2021, Casazza et al., 2019).

A plausible implication of the recent literature is that fusion-space coding now has three interlocking axes: metric optimality through EITFFs and related packings, algebraic construction through harmonic and Radon-Hurwitz methods, and combinatorial generalization through mixed-rank embeddings and design-theoretic families. Taken together, these developments place fusion-space codes at the intersection of frame theory, harmonic analysis, combinatorial design, and high-dimensional coding geometry (Fickus et al., 2024).