---
title: Rotary and Jordan Blocks
url: https://www.emergentmind.com/topics/rotary-and-jordan-blocks
type: topic
---

# Rotary and Jordan Blocks

Rotary and Jordan blocks are foundational structures appearing in both linear algebra and contemporary applications ranging from group theory to machine learning. In matrix analysis, a Jordan block provides the canonical local form for a linear transformation with a single eigenvalue and a single generalized eigenspace. In representation theory and modern neural network architectures, rotary encodings exploit the action of group rotations, while recent advances incorporate non-semisimple (defective) Jordan blocks to realize new kinds of invariant and equivariant mechanisms.

## 1. Jordan Blocks: Algebraic Structure and Canonical Forms

A Jordan block $J_r(\lambda)$ of size $r \times r$ with eigenvalue $\lambda$ is the matrix with $\lambda$ on the diagonal, $1$ on the superdiagonal, and $0$ elsewhere. Any square matrix $A$ over an algebraically closed field can be brought to Jordan canonical form by similarity transformation—decomposing $A$ into a block diagonal matrix of Jordan blocks, one for each eigenvalue. 

For two Jordan blocks $J_r$, $J_s$, the tensor product $J_r \otimes J_s$ decomposes as a direct sum of Jordan blocks, encapsulated by the Jordan partition $\lambda(r, s, p) = (\lambda_1, \ldots, \lambda_r)$, where each $\lambda_i$ is the size of a block, and the sum of the $\lambda_i$ is $rs$ when both input blocks have eigenvalue $1$ over a field of characteristic $p$ [1907.06519]. The block sizes, their multiplicities, and their arrangement encode essential structural information about the joint action of $J_r$ and $J_s$.

## 2. Rotary Operations and Group-Theoretic Encodings

Rotary encodings, particularly Rotary Positional Encoding (RoPE), correspond to semisimple, translation-invariant representations of the additive group $(\mathbb{R}, +)$ on a vector space, where the generator matrix $A$ is diagonalizable (semisimple). The corresponding one-parameter group is $G(d) = \exp(dA)$, and if $A$ has imaginary eigenvalues (e.g., $i\omega$), the action is a planar rotation. In classical RoPE, the $2 \times 2$ blocks are real representations of these rotations, mapping sequence positions to phases and enabling attention mechanisms to be translation-equivariant [2605.04217].

## 3. Jordan Block Decomposition in Representation Theory and Invariant Analysis

Non-semisimple (defective) representations arise when a generator $A$ is not diagonalizable, instead featuring a non-trivial nilpotent part $N$ in a Jordan block. Then, $G(d) = \exp(dJ)$, with $J = \lambda I + N$ and $N^r = 0$ for some $r > 1$. Explicitly, $\exp(dJ) = e^{\lambda d} \sum_{k=0}^{r-1} \frac{d^k N^k}{k!}$. In applications such as positional encoding for transformers, such blocks produce phase responses modulated by polynomial terms in $d$, introducing features like $d e^{i\omega d}$ in addition to pure exponentials. This enables model architectures to express distance-modulated phase structure directly, rather than requiring multiplicative interaction between phase and position [2605.04217].

In mathematical contexts, Jordan block structure governs the action of $p$-groups and the decomposition of homological data. In the topology of angle-valued maps, bar codes (interval decompositions) and Jordan blocks (torsion summands) provide algebraic invariants that relate to Novikov homology, Betti numbers, and monodromy [1303.4328].

## 4. Periodicity, Duality, and Rotary Symmetries in Jordan Block Decompositions

The study of tensor products $J_r \otimes J_s$ over characteristic $p$ fields reveals rich periodic and duality phenomena in the Jordan block decomposition. The multiplicity sequence $c(r, s, p)$ encoding the composition of block sizes is periodic in $s$ with minimal period $p^\beta$ when $r \leq p^\beta$ for some integer $\beta$, according to the Glasby–Praeger–Xia theorem. This periodicity is complemented by a “reflection” property within each period: $c(r, p^\beta + i, p) = \mathrm{rev}(c(r, p^\beta + (p^\beta - i), p))$, where $\mathrm{rev}$ reverses the sequence.

Additional structure, such as subperiodicity and subinterval reflection, emerges for certain $r$ near but not at the boundaries of $p^\beta$. These reflect the underlying combinatorial symmetries of the partitions arising in the Jordan canonical form, which are directly linked to the action of symmetric and dihedral groups [1907.06519].

Further, the so-called “Norman involutions” $\pi(r, s, p)$, permutations naturally associated with the Jordan partition $X(r, s, p)$, precisely encode blockwise reversals and generate the full wreath product $G(r, p) = S_a \wr D_b$ acting on the index set $[r]$, where $a$ is the $p'$-part and $b$ the $p$-part of $r$ [1711.06860].

## 5. Algorithmic and Computational Structures: Zigzag Diagrams, Block Analysis, and Rotary Encodings

Computation of Jordan block decompositions is algorithmically accessible via kernel and cokernel analysis of associated block matrices derived from representations (as in topological data analysis) or by explicit recursion using six-case algorithms for the block composition $c(r, s, p)$ [1303.4328, 1907.06519]. In the context of bar codes and Jordan blocks arising from circle-valued (angle-valued) maps, block decompositions yield both free and torsion invariants—where bar codes correspond to intervals and Jordan blocks to torsion in Novikov-type homology.

For rotary encodings in neural sequence models, the real block realization of non-semisimple (Jordan) representations allows direct encoding of distance-modulated oscillatory features. For example, in Jordan-RoPE, complex $2 \times 2$ Jordan blocks are realized as $4 \times 4$ real blocks with explicit shears that couple rotary and polynomial behaviors, introducing basis elements like $d e^{-\gamma d}\cos(\omega d)$ in a structurally minimal way [2605.04217]. Numerical stabilization is handled via techniques such as bounded shear and scale-normalization, trading off strict group law adherence against dynamic range control.

## 6. Applications and Structural Implications

Jordan and rotary block structures underpin a broad array of applications:

- In algebraic and modular representation theory, the explicit partitioning and periodic/dihedral behavior of Jordan block decompositions classify symmetries and equivariance in module categories and $p$-group actions [1907.06519, 1711.06860].
- In topological data analysis and Morse–Novikov theory, bar codes and Jordan blocks derived from angle-valued maps encode both numerical and torsion invariants, enabling algorithmic computation of refined homological invariants and the direct sum decomposition of homology [1303.4328].
- In neural architectures, (rotary) positional encodings and their Jordan extensions enable models to learn distance-modulated, oscillatory features directly in their primitive attention kernels, increasing expressivity for structured sequence tasks [2605.04217].

The algebraic properties such as periodicity, duality, and wreath-product symmetry realized in the decompositions have clear interpretability in group action, stability, and topological invariance. The non-semisimple Jordan blocks, particularly in group-theoretic or sequence modeling contexts, supply a minimal, yet expressive, foundation for rotational and polynomial invariances and drive complex, yet interpretable, system behaviors.

Source: https://www.emergentmind.com/topics/rotary-and-jordan-blocks