Code Loops: Structured Nonassociative Algebra

• Code loops are highly structured nonassociative algebraic objects derived from doubly even binary codes and characterized by a central extension of an elementary abelian 2-group.
• They are constructed using a code cocycle that satisfies the Griess conditions, ensuring the controlled failure of associativity essential for Moufang loop properties.
• Notable examples like Parker's loop, built from the Golay code, illustrate their significance in the study of sporadic groups and computational algebra.

A code loop is a highly structured algebraic object arising at the intersection of nonassociative algebra and coding theory: specifically, a code loop is a central extension of an elementary abelian 2-group by a 2-element central subgroup, constructed canonically from a doubly even binary code. Introduced by Griess to explain the algebraic pedigree of certain nonassociative loops appearing in the Monster group, code loops have driven substantial advances in the structure theory of Moufang loops, universal algebra, and the interplay between combinatorics and nonassociative systems.

1. Formal Definition and Algebraic Structure

Let $\mathbb{F}_2 = \{0,1\}$ denote the field of two elements, and let $V$ be a finite-dimensional $\mathbb{F}_2$-vector space equipped with a doubly even binary code $\mathcal{C} \subseteq V$, that is, a linear subspace such that $|v| \equiv 0 \pmod{4}$ for every $v \in \mathcal{C}$, where $|v|$ is the Hamming weight. A code loop $L(\mathcal{C})$ is defined on the set $\{\pm 1\} \times \mathcal{C}$ with multiplication given by

$$(\varepsilon, v) \cdot (\delta, w) = (\varepsilon\delta \cdot (-1)^{\theta(v,w)}, v+w)$$

where $$\theta : \mathcal{C} \times \mathcal{C} \to \mathbb{F}_2$$ is a code cocycle satisfying the Griess cocycle conditions: \begin{align*} \theta(v,w) - \theta(u+v,w) + \theta(u,v+w) - \theta(u,v) &\equiv |u\,v\,w| \pmod 2 \ \theta(v,w) + \theta(w,v) &\equiv \frac 12 |v\,w| \pmod 2 \ \theta(v,v) &\equiv \frac 14 |v| \pmod 2 \end{align*} where $|u\,v\,w|$ denotes the size of the coordinatewise product $u \wedge v \wedge w$. Under this multiplication, $L(\mathcal{C})$ is a Moufang loop: it satisfies the Moufang identity and all elements have two-sided inverses, but the operation is nonassociative unless $\mathcal{C}$ is the zero code.

The loop $L(\mathcal{C})$ sits in a central extension: $$1 \longrightarrow \{\pm 1\} \longrightarrow L(\mathcal{C}) \longrightarrow \mathcal{C} \longrightarrow 0$$ The center consists of $\{\pm 1\} \times \{0\}$, and $L(\mathcal{C})/\{\pm 1\} \cong \mathcal{C}$ as elementary abelian $2$-groups. All two-element generated subloops are associative—i.e., they are groups.

2. Construction from Doubly Even Codes and Cocycles

A doubly even binary code $\mathcal{C}$ yields a code loop via the explicit construction above. The central algebraic ingredient is the code cocycle $\theta$, which is tightly constrained by the conditions (CC1)-(CC3) to ensure that the resulting algebra is Moufang, and to control the failure of associativity (i.e., the associator is determined by triple products of codewords). Cocycle normalization ensures that commutator and associator maps are central—that is, for $u,v\in \mathcal{C}$,

$$[u,v] = (-1)^{|u\cap v|/2}, \qquad (u,v,w) = (-1)^{|u\cap v\cap w|}$$

The nontrivial associators force $L(\mathcal{C})$ to be nonassociative but central, an essential property for applications to larger algebraic and geometric constructions (Nagy et al., 2019, Pires et al., 2019).

The prototypical example is Parker's loop, built from the length-24 extended binary Golay code, which plays a central role in the theory of sporadic simple groups and the structure of the Monster group (Nagy et al., 2019).

3. Characteristic Invariants, Classification, and Isomorphism

Each code loop $L(\mathcal{C})$ is uniquely characterized (up to isomorphism) by its squaring map, commutator map, and associator map, and more precisely by its "characteristic vector" comprising the squares and commutators on a set of generators. The paper "Code loops in dimension at most 8" shows that isomorphism classes of code loops on a space $V$ are in bijection with the orbits of a certain affine action of $GL(V)$ on a parameter space determined by the code cocycle and its polarization derivatives (O'Brien et al., 2017).

For small dimensions, the enumeration is explicit:

$\dim V$	Order of $L(\mathcal{C})$	Number of code loops
$6$	$128$	$767$
$7$	$256$	$80,826$
$8$	$512$	$937,791,557$

Classification proceeds stratifying first on the associator map (an alternating trilinear form), then the commutator map, then the squaring map with correction terms as dictated by combinatorial polarization (O'Brien et al., 2017). For ranks 3 and 4, every nonassociative code loop has a unique minimal representation (up to equivalence) by a doubly even code of smallest possible length (Pires et al., 2019, Pires et al., 5 Jan 2026).

4. Variety, Freeness, and Presentations

The class of code loops forms a finitely-based variety $\mathcal{E}$ of Moufang loops, characterized by the identities: $$x^4=1, \;\; [x,y]^2=1, \;\; (x,y,z)^2=1,\; [x^2,y]=1,\; [[x,y],t]=1,\; [(x,y,z),t]=1,$$ and related centrality relations. Free objects in this variety have been constructed explicitly: the free code loop of rank $n$, $F_n$, is the image of the Moufang loop functor $L(G_n)$, where $G_n$ is a nilpotent group of class $3$ with $2n$ generators and triality automorphisms (Pires et al., 5 Jan 2026, Grichkov et al., 2014). Every code loop is a quotient of $F_n$ by a codimension-1 central subgroup not containing the associator.

This provides a universal construction and reduces many structural problems concerning arbitrary code loops to the rank-$3$ case, where explicit presentations and computations are tractable.

5. Representations, Minimality, and Type

Given a code loop $L$ of rank $n$, a representation is a code $V \subseteq \mathbb{F}_2^m$ such that $L \cong L(V)$. There are usually many codes yielding isomorphic code loops, but each nonassociative code loop of small rank (specifically, ranks $3$ and $4$) has a unique minimal reduced representation; reduced means no coordinate class (modulo codeword support) exceeds size $7$. These representations expose the fundamental linkage between the arithmetic of codeword weights and intersection patterns and the algebraic structure of the loop (Pires et al., 2019, Pires et al., 5 Jan 2026).

The minimal degree for a given loop is an invariant, as is the "type": the sorted tuple of sizes of coordinate equivalence classes.

6. Nilpotence, Supernilpotence, and Structural Properties

Code loops belong to the class of Moufang $2$-loops of nilpotency class 2. Every code loop is 3-supernilpotent in the sense of universal algebra: every absorbing polynomial in at least four variables is identically $1$. The variety of $3$-supernilpotent loops admits a finite equational basis, and code loops furnish central examples. Conversely, $k$-supernilpotence does not fully coincide with $k$-nilpotence for loops (unlike groups), but for code loops, $2$-nilpotence implies $3$-supernilpotence, a partial restoration of the group-theoretic collapse (Stanovský et al., 2022).

A key structural property is that all inner mappings are automorphisms and the inner mapping group is abelian (AIM loop). Automorphism and outer automorphism groups for small rank code loops can be explicitly determined and exhibit rich group-theoretic structure, with symmetry groups ranging from $GL_n(2)$ to symmetric and dihedral groups (Grichkov et al., 2014, O'Brien et al., 2017).

7. Computational Aspects and Decomposition

Direct manipulation of code loops, especially for large codes like the extended Golay code (dimension 12), presents computational challenges—the full cocycle table is exponentially large. Recent work demonstrates that, given a decomposition $V = V_1 \oplus V_2$, the full cocycle can be reconstructed from values on the summands and their interactions, reducing the storage and computation required for explicit multiplication to $O(2^{2k})$ for codes of dimension $2k$ (Nagy et al., 2019). This facilitates efficient calculations in loops of order $2^{12}$ and above.

Additionally, code loops may split as direct products over large subspaces where the cocycle vanishes, providing insight into subgroup structure and supporting efficient storage and representation.

---

References:

• (Nagy et al., 2019) (Re)constructing Code Loops
• (Pires et al., 2019) Representations of code loops by binary codes
• (O'Brien et al., 2017) Code loops in dimension at most 8
• (Grichkov et al., 2014) Code loops: automorphisms and representations
• (Pires et al., 5 Jan 2026) Representations of code loops by binary codes
• (Pires et al., 5 Jan 2026) Construction of groups with triality and their corresponding code loops
• (Stanovský et al., 2022) Supernilpotent groups and $3$-supernilpotent loops