Complex Spherical 2-Codes

• Complex spherical 2-codes are finite sets of unit-norm complex vectors with exactly two distinct inner products (complex conjugates), defining a key structure in combinatorial design and optimization.
• Their classification involves algebraic and combinatorial constructions such as doubly regular tournaments and skew Hadamard matrices, which establish sharp bounds on the maximal size relative to dimension.
• Iterative algorithms and semidefinite programming techniques are employed to minimize coherence, paving the way for optimal frame constructions in compressed sensing, wireless communications, and D‑optimal design.

A complex spherical 2-code is a finite set of unit-norm vectors in a complex Euclidean space such that the set of distinct pairwise inner products among the vectors consists of exactly two values, necessarily complex conjugate to each other and not both real. These objects sit at the intersection of finite geometry, algebraic combinatorics, frame theory, and applied disciplines such as compressed sensing and wireless communications. The study of complex spherical 2-codes involves sharp bounds on cardinality, rigorous classification via combinatorial objects such as doubly regular tournaments, skew Hadamard matrices, and supplementary difference sets, and connections to optimal frame constructions and semidefinite programming bounds. Their structural and extremal properties make them central to several optimization and design problems in applied mathematics and engineering.

1. Formal Structure and Absolute Bounds

Let $Q_2(d) = \{x \in \mathbb{C}^d : x^* x = 1\}$ denote the unit sphere in $\mathbb{C}^d$, and let $X \subset Q_2(d)$ be a finite set. The angle set of $X$ is $A(X) = \{x^* y : x, y \in X, x \neq y\}$. A complex spherical 2-code is defined by the property $|A(X)| = 2$ with $A(X) = \{a, \bar a\}$ and $\Im(a) > 0$ (Nozaki et al., 2015). The code size is $n = |X|$; the core theoretical problem is to determine maximal possible $n$ for a given $d$.

The best known absolute bound (“Roy–Suda bound”) is

$$n \leq \begin{cases} 2d + 1, & d \text{ odd} \ 2d, & d \text{ even} \end{cases}$$

Codes that attain this bound are termed tight.

For the further context of codes with prescribed pairwise absolute inner products, the coherence parameter is central: $\mu = \max_{i \neq j} |\langle v_i, v_j \rangle|$ The Welch bound provides a theoretical lower bound for $\mu$ (Zörlein et al., 2014, Conde et al., 2017): $\mu \geq \sqrt{\frac{n-d}{d(n-1)}}$ with equality achieved if and only if the set forms an equiangular tight frame (ETF).

2. Algebraic and Combinatorial Constructions

Tight complex spherical 2-codes have strong links to combinatorial designs:

• For $d$ odd, $n=2d+1$, the adjacency matrix $A$ of the associated tournament must be that of a doubly regular tournament (DRT); equivalently, the set of three matrices $\{I, A, J-A-I\}$ forms a non-symmetric association scheme of class 2 (Nozaki et al., 2015, Roy et al., 2011).
• For $d$ even, $n=2d$, a tight 2-code exists if and only if $I + A - A^{T}$ is a skew-Hadamard matrix. These arise as adjacency structures of tournaments built from such matrices.
• For odd $d$ and $n=2d$, codes are either induced by deleting a vertex from a DRT of order $2d+1$, or by constructing the Seidel matrix from a skew-symmetric D-optimal design (block structure), notably satisfying

$$S^2 \sim_{\text{perm}} \begin{pmatrix} kI + \ell J & kI + \ell J \ kI + \ell J & kI + \ell J \end{pmatrix}$$

for $k, \ell > 0$ integers.

This combinatorial characterization gives a complete classification of tight and extremal (maximal) 2-codes in most cases (Nozaki et al., 2015, Kao et al., 2022). Notably, association schemes arising from these codes are generally non-symmetric, a feature distinct from their real counterparts (Roy et al., 2011).

3. Supplementary Difference Sets and Circulant Constructions

An important class of complex spherical 2-codes comes from skew-symmetric $2-\{v; r, k; \lambda\}$ supplementary difference sets (SDS) in cyclic groups modulo $v$ (Araya et al., 2016). If $(A, B)$ is such a pair, skew-symmetry and difference set equations determine allowable parameter sets. The SDS construction yields a circulant $\pm 1$ block matrix whose structure produces a D-optimal design matrix $M$, which then informs the adjacency matrix of the tournament: $$M := X(R_1, R_2) = \begin{pmatrix} R_1 & R_2 \ -R_2^{T} & R_1^{T} \end{pmatrix}$$ The resulting code $X \subset \mathbb{C}^v$ has $2v$ vectors and pairwise inner products precisely $\{\pm i\sqrt{2(v - 2(r + k - 1))}/v\}$.

Table-based classification is complete for $v \leq 51$, with explicit constructions given for all parameter tuples satisfying the required combinatorial identities and positivity constraints for $r, k, \lambda$ (Araya et al., 2016).

4. Optimization, Coherence Minimization, and Algorithmic Construction

The minimization of code coherence $\mu$ under phase-antipodal equivalence ($v_i \equiv e^{i\phi}v_i$ for all $\phi$) describes the complex line packing problem in the Grassmannian $G(\mathbb{C}^d, 1)$. The problem is recast as searching for vector sets that minimize the maximum absolute inner product up to global phase—a property canonically handled with “Best Complex Antipodal Spherical Codes” (BCASCs) (Zörlein et al., 2014, Conde et al., 2017).

A non-convex potential function is used, with codewords viewed as “charged particles” on the unit sphere: $g(\{s_m\}) = \sum_{m < l} \|s_m - s_l\|^{-(\nu-2)}$ This is driven toward configurations maximizing minimal pairwise Euclidean distance as $\nu \rightarrow \infty$.

The numerical procedure involves an iterative update: $$f_m = \int_0^{2\pi} \sum_{l \neq m} \frac{s_m - e^{i\kappa}s_l}{\|s_m - e^{i\kappa}s_l\|^\nu} d\kappa,\quad s_m \leftarrow \frac{s_m + \alpha f_m}{\|s_m + \alpha f_m\|}$$ with discrete-phase approximations used for efficiency. Fast algorithms rely on phase-sampled approximations and, for large $n$, an approximate nearest neighbor (ANN) search to reduce complexity from $O(n^2)$ per iteration to $O(n \log n)$ (Conde et al., 2017). These approaches yield codes with coherence within a few percent of the Welch bound, asymptotically tightly packing the Grassmannian.

Significant improvements over previous constructions (e.g., Medra–Davidson, Dhillon–Heath–Tropp, Xia–Giannakis) have been numerically demonstrated in various $(d, n)$ regimes (Zörlein et al., 2014).

5. Semidefinite Programming and Upper Bounds

Semidefinite programming (SDP) techniques have supplied the strongest known upper bounds for complex spherical 2-codes (Kao et al., 2022). The method leverages the representation theory of $U(d)$ and its stabilizer $U(d-1)$. Block-diagonalization of moment matrices is achieved via the decomposition of polynomial spaces into harmonic modules with so-called “zonal matrices” $Z^{i, j}_{k, \ell}(x, y)$.

An SDP is formulated using three-point distribution variables $x(u,v,t)$ counting ordered triples with fixed inner products, subject to algebraic constraints and zonal-matrix positivity.

In the 2-code case, the SDP tractably recovers the known absolute bounds, and, in low dimensions, strict upper bounds can be obtained, often outmatching Delsarte linear-programming bounds.

A summary of the key results:

• $|X| \leq 2d+1$, equality iff $d$ odd and adjacency is a DRT;
• $|X| \leq 2d$, equality iff $d$ even and $I+A-A^{T}$ is skew-Hadamard;
• For certain parameters, the SDP yields sharper bounds than linear approaches, particularly when including blocks with $k+\ell > 1$.

6. Applications in Frame Theory and Information Systems

Vectors sets achieving low coherence—particularly BCASCs—play a vital role in frame theory (notably as nearly equiangular tight frames where true ETFs are not available), compressed sensing, and communication systems.

In compressed sensing, the coherence controls the restricted isometry property via the Gershgorin bound, directly impacting sparse recovery guarantees. Empirical evaluation shows that BCASC-based measurement matrices permit higher sparsity and lower oversampling rates than those based on random or Fourier matrices, especially for regimes with small $m/n$ (Zörlein et al., 2014, Conde et al., 2017).

In wireless communications, these constructions inform CDMA signature selection and Grassmannian beamforming codebooks, minimizing worst-case interference by tightly bounding the absolute cross-correlations among channel vectors.

Maximal complex spherical 2-codes with specified algebraic and combinatorial structures also create D-optimal designs and tight configurations with only two Gram matrix eigenvalues, holding significance in statistics and experimental design (Araya et al., 2016).

7. Open Problems and Directions

While the classification of tight and maximal complex spherical 2-codes is comprehensive for moderate dimensions and code lengths, several open directions remain:

• For $v \geq 53$, the exhaustive classification of skew-symmetric $2-\{v; r,k; \lambda\}$ supplementary difference sets, and thus the existence of corresponding codes, is unsettled. Infinite families are conjectured based on Paley constructions for $p \equiv 3 \bmod 4$ (Araya et al., 2016).
• The uniqueness of SDS constructions for larger $v$ and the possibility of sporadic, non-paley codes remain open.
• The behavior and optimality of BCASC algorithms in high-dimensional regimes and their asymptotic coherence gaps require further investigation.
• The development of SDP hierarchies beyond the initial harmonic blocks may provide sharper or even exact bounds in yet-unexplored parameter regions (Kao et al., 2022).

Complex spherical 2-codes thus provide a rich mathematical framework with deep connections to combinatorial design, semidefinite optimization, theoretical and applied information theory, and signal processing.