Permutation and Orthogonal Equivariance

• Permutation and orthogonal equivariance are invariance properties under symmetric and orthogonal group actions, ensuring canonical matrix forms and robust optimization frameworks.
• They reveal deep symmetries in structures such as orthogonal arrays, antisymmetric matrices, and ILP formulations, which enhance classification and algorithmic efficiency.
• Constructive algorithms exploiting these equivariances yield stable decompositions that reduce computational complexity in spectral analysis and experimental design.

Permutation and orthogonal equivariance are fundamental notions in matrix theory, optimization, and combinatorial design, describing invariance properties of structures and algorithms under group actions. Permutation equivariance typically refers to invariance under actions of the symmetric group, such as relabeling indices, while orthogonal equivariance generalizes this to invariance under transformations from the orthogonal group, preserving inner products and symmetries associated with matrix forms. The study of these equivariances elucidates deeply rooted symmetries in integer linear programming (ILP), orthogonal arrays (OA), and the spectral decomposition of antidiagonal operators, with significant consequences for combinatorial optimization, experimental design, and linear algebra.

1. Permutation Equivariance: Definition and Canonical Examples

Permutation equivariance describes the invariance of mathematical objects under permutations, typically encoded by the symmetric group $S_n$. For matrices or arrays, permutation equivariance is manifested via permutation similarity: for $A \in \mathbb{C}^{n \times n}$ and $P$ a permutation matrix ($P \in S_n$), the map $A \mapsto P^T A P$ realizes permutation similarity. In optimization, permutation equivariance allows relabeling of variables or indices without impacting feasible sets or objective values. The canonical case, detailed in "The linear programming relaxation permutation symmetry group of an orthogonal array defining integer linear program" (Arquette et al., 2021), considers the symmetry group $S_s \wr S_k = S_s^k \rtimes S_k$ for $k$-factor, $s$-level orthogonal arrays, where $S_s^k$ independently permutes levels and $S_k$ permutes factors, reflecting invariances in the OA-ILP constraints and solution space. This symmetry group is embedded naturally in the LP relaxation symmetry group, and equivariance under $S_s \wr S_k$ underpins efficient optimization and enumeration.

2. Orthogonal Equivariance and Extensions

Orthogonal equivariance refers to invariance under similarity transformations by orthogonal matrices ($Q \in O(n)$), i.e., $M \mapsto Q^T M Q$. This concept generalizes permutation equivariance beyond discrete symmetries to continuous group actions, preserving the spectral and geometric properties of the operator. In the context of antidiagonal and antisymmetric matrices, orthogonal equivariance respects the underlying structure and commutes with the Schur decomposition. As detailed in "Antidiagonal Operators, Antidiagonalization, Hollow Quasidiagonalization -- Unitary, Orthogonal, Permutation, and Otherwise - and Symmetric Spectra" (Nicholus, 2023), real antisymmetric matrices are a single orbit under orthogonal similarity, and orthogonal antidiagonalization delivers canonical hollow-quasidiagonal forms, preserving eigenvalues and spectral attributes. This provides both a matrix-theoretic version of equivariance and a computational route for spectral decomposition.

3. Permutation Symmetry Groups in Orthogonal Array ILP and LP Relaxations

For the ILP formulation of orthogonal arrays OA($N,k,s,t$), symmetry groups dictate the possible relabelings of array runs while preserving equality constraints and feasibility. The permutation symmetry group of the LP relaxation, denoted $G^{LP}$, is given by $$G^{LP} = \operatorname{Aut}(\text{Row}(M)) \cap S_{s^k}$$, where $M$ is the constraint matrix. In the $2$-level (binary) case, with strength $t=1$, it is proved that $G^{LP} \cong S_2 \wr S_k$, corresponding precisely to all signed permutations of columns and indices. With strength $t=2$, for $k \geq 4$, a previously unknown lift occurs: $G^{LP} \cong S_2^k \rtimes S_{k+1}$, with $S_{k+1}$ permuting main effects and "all-but-one-factor" directions, revealing hidden symmetries ("orthogonal equivariance") beyond mere permutation of indices (Arquette et al., 2021). This deeper symmetry informs both theoretical understanding and algorithmic applications.

4. Equivariant Decompositions: Permutation, Orthogonal, and Unitary Similarity

Three principal decompositions correspond to different group-equivariant structures:

Decomposition type	Equivariance	Normal form
Permutation–similarity	$A \mapsto P^TAP, P \in S_n$	Hollow-quasidiagonal $\bigoplus 2 \times 2$ blocks
General similarity (Jordan)	$M \mapsto V^{-1} M V$	Direct sum of Jordan blocks ($J_2(0)$, $[\lambda]$)
Unitary similarity (Schur)	$M \mapsto U^* M U, U \in U(n)$	(Upper) Schur/quasidiagonal (diagonal or $2 \times 2$) blocks

Permutation equivariance yields a canonical hollow-quasidiagonal normal form for traceless antidiagonalizable matrices. Orthogonal equivariance extends this to the real Schur decomposition for real antisymmetric matrices, aligning antidiagonalization with the $O(n)$ symmetry group. General similarity and unitary similarity correspond, respectively, to Jordan and Schur decompositions, each delivering a canonical block-diagonal form that reflects the underlying group symmetry (Nicholus, 2023). These decompositions facilitate classification, spectral analysis, and algorithmic diagonalization.

5. Constructive Algorithms and Computational Implications

Explicit algorithms have been developed for both permutation and orthogonal equivariant normal forms. For permutation quasidiagonalization, antidiagonal entries are paired and mapped via a fixed permutation to $2 \times 2$ hollow blocks, independent of magnitude or sign. For real Schur and orthogonal antidiagonalization, any real antisymmetric matrix can be decomposed by first applying a standard QR-based real Schur decomposition and then a fixed permutation, yielding an antisymmetric antidiagonal matrix whose nonzero entries are the eigenvalues. All steps are numerically stable and group-equivariant (Nicholus, 2023). In OA-ILP, leveraging permutation and orthogonal symmetries enables large-scale pruning of isomorphic nodes, dramatically reducing computational complexity for highly symmetric experimental designs (Arquette et al., 2021).

6. Geometric and Combinatorial Significance of Permutation and Orthogonal Equivariance

The emergence of hidden symmetries—especially the lift from $S_s \wr S_k$ to $S_2^k \rtimes S_{k+1}$ at even strengths—indicates that combinatorial orthogonality constraints endow OA-polytopes with automorphisms beyond simple row/column relabelings. A main-effect direction can be mapped to averaged interaction directions, consistent with system constraints only when coordinated by the larger $S_{k+1}$ action, reflecting new "orthogonal equivariance" phenomena. This suggests that the full symmetry group combines discrete permutation invariance and deeper orthogonal-like interchanges, shaping the structure, enumeration, and optimization in OA polytopes and antidiagonal operator theory. Such symmetries both deepen the understanding of combinatorial polytopes and endow powerful tools for classification and algorithmic efficiency (Arquette et al., 2021, Nicholus, 2023).

7. Concluding Perspectives and Open Problems

Permutation and orthogonal equivariance are key for both theoretical symmetry classification and practical reduction of computational effort. Current evidence, including empirical results for small cases and constructive decompositions, suggests the following conjectural status for the OA-ILP formulation: for $s=2$, $t$ odd, $G^{LP} \cong S_2 \wr S_k$; for $s=2$, $t$ even, $G^{LP} \cong S_2^k \rtimes S_{k+1}$; for $s > 2$, all $t$, $G^{LP} \cong S_s \wr S_k$ (Arquette et al., 2021). Further investigation into orthogonal-like lifts and the full automorphism group of OA-polytopes remains an open direction, with potential implications for combinatorial design, algebraic statistics, and large-scale optimization.