Higher-Order Orthogonality

• Higher-order orthogonality is a framework that extends pairwise orthogonality to multidimensional objects like tensors, polynomials, and arrays using advanced algebraic, combinatorial, and analytic identities.
• It underpins methods such as ODECO/UDECO, HOSVD, and HOOI by ensuring unique decompositions and convergence guarantees through structured orthogonality constraints.
• The concept has broad applications across spectral theory, harmonic analysis, and combinatorial designs, impacting fields from quantum mechanics to signal processing.

Higher-order orthogonality encompasses generalizations of classical orthogonality from functions, vectors, and matrices to higher-rank tensors, arrays, polynomials, and combinatorial or analytic structures. The subject spans multilinear algebra, special function theory, harmonic and Fourier analysis, combinatorics, and representation theory, and is characterized by forms of orthogonality or mutual independence among higher arity objects—modeled via symmetry, algebraic identities, or vanishing of certain mixed inner products—beyond the scope of ordinary pairwise orthogonality.

1. Algebraic Structures: Orthogonally Decomposable Tensors

The fundamental problem in the algebraic context concerns tensors which admit orthogonal or unitary decompositions (ODECO/UDECO). If $K=\mathbb{R}$, a tensor $T\in V_1\otimes\cdots\otimes V_d$ is called orthogonally decomposable (ODECO) if $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$ with the vectors in each mode pairwise orthogonal. The complex (unitary) case yields the UDECO notion (Boralevi et al., 2015). For symmetric and alternating tensors, analogous notions apply, with decompositions into symmetric or antisymmetric rank-one summands subject to pairwise orthogonality constraints.

ODECO/UDECO tensors for $d\geq 3$ have unique decompositions, and the sets of such tensors form real-algebraic varieties cut out by polynomial identities:

• Ordinary real: degree-2 partial associativity.
• Ordinary complex: degree-3 partial semi-associativity.
• Symmetric real: degree-2 commutative associativity.
• Symmetric complex: degree-3 semi-associativity.
• Alternating real: degree-2 Jacobi and degree-4 cross identities.
• Alternating complex: degree-3 Casimir plus degree-4 cross identity.

The flattening argument shows the full higher-order varieties are characterized by those of order 3, demonstrating the algebraic and geometric unity of higher-order orthogonality varieties (Boralevi et al., 2015).

2. Higher-Order Orthogonality in Tensor Decompositions and HOSVD

In multilinear algebra, higher-order orthogonality arises in tensor decompositions such as the Higher-Order Singular Value Decomposition (HOSVD) and Higher-Order Orthogonal Iteration (HOOI).

• HOSVD: Any third-order tensor $\mathcal{T}\in\mathbb{C}^{I_1\times I_2\times I_3}$ admits a decomposition $$\mathcal{T} = (U^{(1)}\otimes U^{(2)}\otimes U^{(3)})\cdot \mathcal{S}$$, where each $U^{(n)}$ is unitary and the core tensor $\mathcal{S}$ satisfies all-orthogonality: fixing the $n$-th index, the resulting subtensors are mutually orthogonal (Choong et al., 2020). This guarantees that the mode-$n$ unfoldings produce diagonal one-body reduced density matrices when viewed as quantum states, linking multilinear algebra to quantum entanglement structures.
• HOOI: The HOOI algorithm seeks the best low multilinear-rank approximation of a tensor and iteratively projects onto Stiefel manifolds enforcing orthonormality. Block-nondegeneracy—strict singular value gaps—ensures uniqueness of factor subspaces. Recent results show that, under block-nondegeneracy, both greedy and standard HOOI converge in the subspace sense, settling a longstanding question (Xu, 2015).

3. Special Function Theory: Higher-Order Orthogonality in Polynomials

Higher-order orthogonality is a key property of orthogonal polynomial families defined by higher-order differential or difference equations and nonclassical measures.

• Generalized Jacobi Polynomials: Koornwinder’s polynomials $T\in V_1\otimes\cdots\otimes V_d$0 are orthogonal with respect to a measure that includes both a classical Jacobi weight and endpoint masses. Their eigenvalue problem involves a differential operator of order $T\in V_1\otimes\cdots\otimes V_d$1 expressible as a sum of component operators, each symmetric with respect to the associated inner product. Orthogonality of the polynomials encompasses both the continuous and discrete components—a genuinely higher-order orthogonality (Markett, 2017, Markett, 2017).
• Higher-Order Sobolev-Type Orthogonality: Sobolev-type polynomials, such as $T\in V_1\otimes\cdots\otimes V_d$2, are orthogonal with respect to inner products involving both $T\in V_1\otimes\cdots\otimes V_d$3-integrals and discrete masses at boundary $T\in V_1\otimes\cdots\otimes V_d$4-derivatives of arbitrary order. The resulting polynomials satisfy holonomic q-difference equations of higher order, possess nonstandard three-term recurrence relations, and give rise to rational J-fraction expansions—reflecting the depth of higher-order orthogonality in the moment problem and analytic function theory (Hermoso et al., 2020).

4. Higher-Order Orthogonality in Harmonic/Fourier Analysis

Higher-order orthogonality plays a central role in modern Fourier analytic and ergodic-theoretic approaches to number theory.

• Gowers Uniformity Norms and Aperiodic Multiplicative Functions: A sequence $T\in V_1\otimes\cdots\otimes V_d$5 possesses higher-order orthogonality if its (normalized) Gowers uniformity norms $T\in V_1\otimes\cdots\otimes V_d$6 vanish as $T\in V_1\otimes\cdots\otimes V_d$7. Structure theorems show that $T\in V_1\otimes\cdots\otimes V_d$8 can be decomposed into an approximately periodic component and a component with small $T\in V_1\otimes\cdots\otimes V_d$9-norm (Frantzikinakis et al., 2014). A multiplicative function is aperiodic if and only if all its Gowers norms vanish (for all $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$0), capturing an intrinsic higher-order independence.
• Orthogonality to Nilsequences: Aperiodic multiplicative functions are asymptotically orthogonal to all totally equidistributed polynomial nilsequences, generalizing Daboussi-Delange orthogonality to irrational sequences in nonabelian settings (Frantzikinakis et al., 2014). This informs zero-correlation conjectures such as the Chowla conjecture in the context of higher-degree polynomials.

5. Combinatorial and Array-Theoretic Generalizations

Higher-order orthogonality is foundational in combinatorial design and multi-dimensional array theory, especially for perfect autocorrelation and complementary constructions in radar, communications, and coding theory.

• Generalized Array Orthogonality Property (GAOP): An $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$1-dimensional array $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$2 (with values in roots of unity) is said to have GAOP for divisor $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$3 if its associated $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$4-dimensional array $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$5 has the property that all $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$6 slices are pairwise orthogonal and form a periodic complementary set (Blake et al., 2014). GAOP is a precise multidimensional abstraction, guaranteeing that the original $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$7 is a perfect array—all off-peak shifts in periodic autocorrelation vanish. This encapsulates and generalizes the classical array orthogonality property in 1D sequence theory.

Key constructions (such as generalized Frank or Milewski-Chu arrays) leverage GAOP, and its proof integrates combinatorial reindexing and character sum identities. The property underpins applications in radar and wireless systems requiring higher-dimensional ambiguity function control (Blake et al., 2014).

6. Representation-Theoretic and Spectral Higher-Order Orthogonality

In the context of automorphic forms and representation theory, higher-order orthogonality is captured by explicit relations among Fourier coefficients of cusp forms for higher rank groups, with implications for spectral statistics and arithmetic applications.

• GL$$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$8 Orthogonality: Explicit orthogonality relations for Fourier-Whittaker coefficients of automorphic cusp forms on $$T = \sum_{i=1}^k v_{i1}\otimes v_{i2}\otimes\cdots\otimes v_{id}$$9, such as the GL(4) orthogonality relation, embody higher-order orthogonality at the spectral level (Goldfeld et al., 2019). These results are obtained via the Kuznetsov trace formula and involve Weyl group summations, Kloosterman sum bounds, and multi-variable harmonic analysis. Power-saving error terms, new polynomial test functions, and geometric-spectral dualities are principal technical achievements. Analogous phenomena underlie GL(2) and GL(3) cases but with increasing complexity for higher $d\geq 3$0 (Goldfeld et al., 2019).

These spectral results facilitate averaging identities, subconvex bounds for L-functions, distribution of Fourier coefficients, and reinforce symmetry type predictions for families of L-functions.

7. Synthesis: Principles and Cross-Disciplinary Influence

Across these domains, higher-order orthogonality is characterized by:

• Mutual orthogonality of multiple objects beyond classical pairwise settings, enforced algebraically (identities, polynomial vanishing), analytically (multi-norms, spectral decompositions), or combinatorially.
• The conceptual unification of independence phenomena—be it for tensors, polynomials, multiplicative functions, or array structures—via higher-order orthogonality serves as a touchstone for theory and applications.
• Interaction with spectral theory, noncommutative harmonic analysis, operator theory, and combinatorics elucidates general mechanisms of signal separation, randomness, and structural rigidity in systems with complex symmetries or constraints.

The above frameworks continue to inform contemporary advances in algebraic geometry, quantum information, harmonic analysis, and theoretical computer science.