Convexity of Neural Codes

• Convex neural codes are defined as collection of codewords arising from intersections of convex open sets in Euclidean space, highlighting minimal embedding dimensions.
• Major classes include max intersection-complete, inductively pierced, and polytope-convex codes, each characterized by unique geometric and combinatorial conditions.
• Recognition involves analyzing topological obstructions and algebraic invariants, with computational challenges linked to NP-hardness and ER-completeness in identifying convex realizability.

A neural code is convex if it arises as the set of intersection patterns of convex open sets in Euclidean space, specifically, if for each codeword $c \subseteq [n]$ there is a region of the form $$\bigcap_{i\in c} U_i \setminus \bigcup_{j\notin c} U_j$$ that is nonempty for some collection of convex open sets $(U_1, \ldots, U_n)$ in $\mathbb{R}^d$. The study of convexity of neural codes seeks to characterize, for a given combinatorial code $\mathcal{C}$, whether such a realization exists, what minimal dimensionality is required, and which geometric, combinatorial, and algebraic invariants or constructions govern the answer.

1. Definition and Fundamental Criteria

A neural code $\mathcal{C}$ on $n$ neurons is convex if there exists a family of convex open sets $U_1, \ldots, U_n$ in $\mathbb{R}^d$ such that $\mathcal{C}$ is precisely the set of subsets $c\subseteq [n]$ for which

$$\bigcap_{i \in c} U_i \setminus \bigcup_{j \notin c} U_j \ne \emptyset.$$

The minimal such $d$—the embedding dimension—reflects how complex or high-dimensional an environment is needed to realize the code combinatorially.

Not every neural code is convex. Initial necessary conditions for convexity are captured by the absence of local obstructions: if a codeword $\sigma\in\Delta(\mathcal{C})\setminus\mathcal{C}$ (where $\Delta(\mathcal{C})$ is the simplicial complex generated by the code) is an intersection of maximal codewords with non-contractible link, then convexity fails. This criterion is both necessary and sufficient for convex codes with up to three maximal codewords (and for all codes on at most four neurons), but not in general for codes with more maximal codewords (Lienkaemper et al., 2015, Johnston et al., 2020, Ahmed et al., 23 Oct 2025).

2. Major Classes of Convex Neural Codes

Max Intersection-Complete and Strongly Max-Intersection-Complete Codes

A code is max intersection-complete if it contains all nonempty intersections of maximal codewords. Such codes are always convex, and both open and closed convex (Cruz et al., 2016). Strongly max intersection-completeness (requiring that every intersection of a codeword with any number of maximal codewords is present) further strengthens this guarantee, with dimension bounds that improve upon the general case when the intersection graph has grid-like structure (Williams, 2016).

Inductively Pierced Codes

Inductively $k$-pierced codes form a constructible subclass characterized by an explicit iterative "piercing" operation. Any inductively $k$-pierced code is nondegenerate convex (preserved under small perturbations) and has a ball realization in $\mathbb{R}^{k+1}$, as well as a nondegenerate hyperplane realization in $\mathbb{R}^n$ (Lienkaemper, 2018). The class sits inside the set of nondegenerate convex codes and further guarantees additional strong combinatorial properties, such as shellability of the associated polar complex.

Polytope Convexity and Oriented Matroids

A code that is the image of a representable oriented matroid code under a code morphism is said to be polytope-convex or combinatorially convex. Every code with a realization by interiors of convex polytopes (as opposed to arbitrary convex open sets) is a minor of a representable oriented matroid code. In the plane, this characterizes all open convex codes; in higher dimensions, the inclusion may be strict (Kunin et al., 2020). The connection is categorical, with functorial relationships between oriented matroids, neural codes, and their associated rings.

3. Obstructions and Recognition Complexity

Local and Global Obstructions

Local obstructions, defined via the topology of links in the code's simplicial complex, preclude convexity but do not characterize it in general: explicit counterexamples for codes on five or more neurons (or four or more maximal codewords) exist (Lienkaemper et al., 2015, Ahmed et al., 23 Oct 2025). Global combinatorial phenomena, such as the order-forcing property and configuration-based obstructions like "wheels," also play a role (Jeffs et al., 2020, Ahmed et al., 23 Oct 2025). For most codes with four or fewer maximal codewords, convexity can be characterized by the absence of local obstructions (and wheels where necessary), but the situation grows more complex as the number increases.

Algebraic Criteria

The neural ideal and its canonical form provide algebraic signatures of convexity and non-convexity. In particular, algebraic relations among codewords (pseudo-monomials in the canonical form) correspond to forbidden intersection patterns or covering relationships. Some signatures detect all known local obstructions, but not all non-convex codes are currently detectable this way (Curto et al., 2018).

Topological and Combinatorial Structures

For codes realizing stable hyperplane arrangements, the polar complex—encapsulating the code's bitflip-invariant information—is always shellable, and shellability eliminates all known obstructions to such realizations (Itskov et al., 2018). Simplicial complexes of certain convex codes exhibit further properties, such as being vertex-decomposable or contractible, contributing to their structural robustness (Lienkaemper, 2018).

Complexity

The recognition problem for convex neural codes is computationally intractable: deciding whether a code is convex is as hard as the representability of (uniform, rank-3) oriented matroids—an $\exists\mathbb{R}$-complete (i.e., ER-complete) and NP-hard problem (Kunin et al., 2020, Lienkaemper, 2022). This precludes any hope for a finitary combinatorial characterization of convex codes in general.

4. Relationships Between Open and Closed Convexity

Open convex and closed convex codes do not coincide: there exist codes which are open convex but not closed convex, and vice versa (Cruz et al., 2016, Gambacini et al., 2019, Chan et al., 2020). The distinction arises primarily for degenerate codes. The nondegeneracy condition (preservation of codewords under closure/interior) ensures the equivalence of open and closed convex realizations. For nondegenerate codes, the minimal dimensions for open and closed realizations agree (Chan et al., 2020).

Monotonicity under codeword addition—a desirable property where adding non-maximal codewords preserves convexity—holds for open convex codes but fails for closed convex codes. Moreover, adding codewords can increase the minimal embedding dimension for closed convex codes arbitrarily, while for open convex codes, the increase is at most one (Gambacini et al., 2019).

5. Structural and Algorithmic Insights

Morphisms, Trunks, and Poset Structure

Morphisms of neural codes, defined via preimages of trunks (collections of codewords containing a specified set of neurons), preserve convexity. The partial order on codes induced by morphisms and trunk operations organizes the universe of neural codes; convex codes form a down-set. Minimally non-convex codes—those for which every morphic image and trunk is convex—serve as atomic obstructions whose understanding could lead to a full classification of convex codes (Jeffs, 2018).

Neural ring homomorphisms correspond to meaningful code maps in the algebraic setting, restricting to those transformations under which convexity and minimal embedding dimension are preserved or decrease (Curto et al., 2015).

Dimension Bounds and Constructions

Sharp dimension bounds are known for several large classes. Every $k$-inductively pierced code is convex in dimension at most $k+1$ by explicit geometric construction. Strongly max intersection-complete codes with certain grid-like intersection graphs are guaranteed to be convex in dimension at most $d+2$, where $d$ is the dimension of the grid (Williams, 2016, Lienkaemper, 2018). For $2$-sparse codes, convex realizability is completely characterized by intersection completeness, and minimal embedding dimension is at most $3$ (Jeffs et al., 2015).

However, in the absence of the open/closed restriction, every code is convex in high enough dimension ($\mathbb{R}^{k-1}$, for $k$ nonempty codewords), so openness/closedness is essential for the mathematical and biological relevance of convex codes (Franke et al., 2017).

Tables

Class of code	Convexity Criterion	Minimal dimension
Max intersection-complete	All intersections of maximals present	$\leq \max\{2,k-1\}$
Strongly max intersection-complete + suitable graph	SMIC + grid-like $G_C$	$\leq d+2$
$k$-inductively pierced	Construction via $k$-piercings	$\leq k+1$
2-sparse	Intersection-complete supports	$\leq 3$

6. Broader Implications and Open Directions

The landscape of convex neural codes is shaped by a delicate interplay between combinatorics, geometry, algebra, and computational complexity. Convex code theory provides explicit, robust families of convex neural codes with low embedding dimension, stable realizability, and algebraic/topological invariants, offering models for plausible neural representations in sensory and hippocampal circuits. The field continues to advance through the identification of new obstructions and structures (e.g., wheel obstructions, non-local combinatorial orderings), refined algebraic and topological signatures, and new connections to oriented matroids and convex geometry.

A central open problem remains: to further classify minimally non-convex codes and to better understand the full suite of combinatorial, algebraic, and topological invariants that govern convex realizability, particularly as neuron number and code complexity grow. Recognition problems also remain computationally challenging, motivating research on efficient invariants and heuristic algorithms for practical code identification in mathematical neuroscience.