Directed Flag Complex

• Directed flag complex is an ordered simplicial complex formed by promoting transitive directed cliques in a digraph, preserving inherent asymmetry.
• It employs on-the-fly simplex generation with branch pruning techniques, as exemplified by the Flagser software, to compute homology efficiently.
• Its applications in neuroscience and molecular network analysis reveal rich homotopy types and directional spectral features not seen in undirected models.

Searching arXiv for recent and foundational papers on directed flag complexes to ground the article. arXiv search query: "directed flag complex"

A directed flag complex is the ordered simplicial complex attached to a directed graph by promoting each transitively oriented directed clique to a simplex. Concretely, for a digraph $G=(V,E)$, a $k$-simplex is an ordered $(k+1)$-tuple $(v_0,\dots,v_k)$ of distinct vertices such that $v_i\to v_j$ for all $i<j$. This construction is the directed analogue of the ordinary flag, or clique, complex, but it retains asymmetry that would be erased by passing to the underlying undirected graph. It has become a central object in computational topology on directed networks, with applications ranging from neuronal microcircuitry to persistent spectral methods on molecular interaction graphs (Luetgehetmann et al., 2019, Jones et al., 2023).

1. Definition and combinatorial character

The defining feature of a directed flag complex is that simplices are ordered. In the undirected case, a simplex is determined by a vertex set; in the directed case, the order is part of the data. The standard definition is

$$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$

Thus every earlier vertex must point to every later vertex. The vertices $v_0$ and $v_k$ are naturally interpreted as the source and sink of the simplex (Luetgehetmann et al., 2019).

Several immediate consequences distinguish directed flag complexes from ordinary clique complexes. A directed simplex is not an arbitrary cyclic pattern: the directed 3-cycle

$(w_0,w_1),\quad (w_1,w_2),\quad (w_2,w_0)$

does not define a directed $k$0-simplex, because the required transitive edge $k$1 is missing relative to the order $k$2. More generally, a directed flag complex records transitively oriented complete subgraphs, not merely strongly connected or cyclic ones. This is why every face of a directed simplex is again a directed simplex, so $k$3 is an ordered simplicial complex, whereas not every ordered simplicial complex is a directed flag complex (Zia et al., 2024).

The same underlying vertex set can support more than one simplex when reciprocal edges are present. In the complete bidirected graph on $k$4 vertices, every ordering of the vertex set determines a directed $k$5-simplex, so the complex contains $k$6 top-dimensional simplices on the same set of vertices; in that special case, the complex is the complex of injective words on $k$7 letters. By contrast, if $k$8 is a tournament, directed cliques are exactly transitive subtournaments, and each simplex is then determined uniquely by its underlying vertex set. This is a special case in which the ordered structure collapses back to an ordinary simplicial-complex description (Govc, 2020).

2. Algebraic, functorial, and directional structure

Once $k$9 is formed, its chain theory uses the standard alternating simplicial boundary on ordered simplices: $(k+1)$0 Accordingly,

$(k+1)$1

A useful structural fact is that when the directed flag complex is viewed through regular path-complex formalism, no extra correction is needed: for ordered simplicial complexes one has $(k+1)$2, so the directed-flag-complex chain complex is exactly the ordinary chain complex on ordered directed cliques. At the categorical level, $(k+1)$3 is functorial on the class of triangle-collapsing digraph maps, and a homotopy-like relation on such maps induces identical maps on directed-flag-complex homology; the same framework yields an equivalence relation on digraphs with isomorphic directed-flag-complex homology (Chaplin et al., 2024).

A recurrent limitation of ordinary homology and of the geometric realization is that they may forget much of the directional organization present combinatorially. This motivated the directed extension of Atkin’s $(k+1)$4-connectivity: simplices are related not merely by sharing a $(k+1)$5-face, but by sharing it through prescribed directional face maps $(k+1)$6. The resulting $(k+1)$7-connectivity relation is a preorder rather than an equivalence relation, and via Alexandroff correspondence it defines finite topological spaces whose homotopy types can differ even when the underlying directed flag complexes have the same ordinary homology. This yields direction-sensitive notions such as simplicial flow and longest simplicial paths in directed networks (Riihimäki, 2022).

3. Computation and software

The principal software system for large-scale computation is Flagser, which was designed specifically to construct directed flag complexes and compute their homology and persistent homology efficiently (Luetgehetmann et al., 2019). Its core design choice is to avoid storing the entire complex whenever possible. Instead, simplices are generated on the fly from the digraph.

The enumeration strategy exploits the fact that every simplex has a unique initial vertex. Fixing $(k+1)$8, one extends a simplex $(k+1)$9 by intersecting outgoing-neighbor sets: $(v_0,\dots,v_k)$0 This supports branch pruning, since failure of the clique condition at one stage prevents all further cofaces, and it parallelizes naturally by distributing initial vertices across processors. To accelerate intersections, Flagser stores the adjacency matrix and its transpose as bitsets, reducing memory by about a factor of $(v_0,\dots,v_k)$1 relative to byte storage and turning set intersection into bitwise logical AND. For cohomology computations, it precomputes sparse coboundary matrices rather than relying on the indexing tricks used by Ripser, since those tricks do not scale well in the directed setting (Luetgehetmann et al., 2019).

Flagser also supports filtrations derived from weights on vertices or edges and extended to higher simplices, for example by assigning to each simplex the maximum weight of an edge it contains. It includes an Approximate mode in which columns that exceed a prescribed reduction-step threshold are skipped; the theoretical error is bounded by the number of skipped columns. A notable warning is that Ripser’s “apparent pairs” optimization cannot safely be used for general directed flag complexes, because it can return incorrect Betti numbers in directed examples (Luetgehetmann et al., 2019).

The scalability can be extreme. The paper reports construction of the directed flag complex of a Blue Brain Project neocortical column graph with roughly $(v_0,\dots,v_k)$2K vertices and $(v_0,\dots,v_k)$3M directed edges in 23.76s using 1.08GB of memory, and much larger simplex-count computations on HPC systems for datasets with millions of vertices and billions of edges (Luetgehetmann et al., 2019).

4. Topological results and homotopy types

The directed flag complex is topologically much richer than the undirected clique complex obtained by forgetting orientation. The most detailed case study to date concerns the chemical synapse network of C. elegans. Its directed flag complex is $(v_0,\dots,v_k)$4-dimensional, has simplex counts

$(v_0,\dots,v_k)$5

in dimensions $(v_0,\dots,v_k)$6 through $(v_0,\dots,v_k)$7, and Betti numbers

$(v_0,\dots,v_k)$8

Its integral homology is torsion-free, and the complex was shown to be homotopy equivalent to a wedge of spheres,

$(v_0,\dots,v_k)$9

The corresponding undirected flag complex has nontrivial homology only up to degree $v_i\to v_j$0, which makes the role of directionality explicit rather than cosmetic (Govc, 2020).

Tournament examples expose a different regime. For every tournament $v_i\to v_j$1, the fundamental group $v_i\to v_j$2 is free. All directed flag complexes of tournaments on $v_i\to v_j$3 vertices are wedges of spheres, but larger regular and doubly regular tournaments exhibit wedges of Moore spaces, including torsion phenomena such as $v_i\to v_j$4, $v_i\to v_j$5, and $v_i\to v_j$6. The paper also distinguishes the larger flag tournaplex, whose simplices are all tournaments contained in the digraph, not only the transitive ones used by $v_i\to v_j$7; in the C. elegans case, stages of the local directionality filtration on the flag tournaplex exhibit $v_i\to v_j$8-torsion even though the directed flag complex itself is torsion-free (Govc, 2020).

5. Persistent and spectral extensions

The directed flag complex supports a Hodge-type Laplacian theory. For the boundary matrices $v_i\to v_j$9 of $i<j$0, the directed flag Laplacian is

$i<j$1

These operators are self-adjoint and positive semidefinite, and the multiplicity of the zero eigenvalue equals the Betti number $i<j$2. The nonzero spectrum carries additional geometric information not captured by homology alone (Jones et al., 2023).

Persistent directed flag Laplacians refine this to filtered digraphs. Given a filtration $i<j$3 of directed flag complexes and levels $i<j$4, one defines

$i<j$5

the persistent boundary

$i<j$6

and then

$i<j$7

The harmonic spectrum recovers persistent Betti numbers: $i<j$8 A basic example shows why the nonzero spectrum matters: adding a single edge can create a $i<j$9-simplex and a $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$0-cycle simultaneously, with no change in persistent homology because the cycle is immediately filled, yet the smallest nonzero persistent Laplacian eigenvalue still changes (Jones et al., 2023).

An application-driven version of this theory is the persistent directed flag Laplacian model for protein–ligand binding affinity prediction. There the underlying digraph is bipartite between protein and ligand atoms; edge directions are assigned by electronegativity comparison, weights come from transformed FRI kernel values $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$1, and the filtration scans $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$2 in increments of $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$3. The directed flag complexes generated across the filtration support PDFL spectra in dimensions $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$4 and $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$5; ten statistics are extracted from eigenvalues over five filtration bins and $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$6 element-specific interaction types, yielding $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$7 features per complex in the one-kernel model. The reported result is that the multi-kernel PDFL model outperforms competing methods on PDBbind v2007, v2013, and v2016, which suggests that directed higher-order interaction motifs can be predictive when the asymmetry is physically meaningful (Zia et al., 2024).

6. Applications, related constructions, and limitations

Directed flag complexes have been used in at least three distinct ways: as topological summaries of directed biological networks, as substrates for spectral-topological descriptors, and as objects against which null models can test whether observed homology is statistically exceptional. In the null-model setting, a recent MCMC framework samples directed graphs with the same underlying undirected graph and approximately the same simplex-count vector as the observed graph. The state space is explored by Single Edge Flip, Double Edge Move, Clique Permute, and Clique Swap moves, while simplex counts are updated locally. Applied to C. elegans and Blue Brain Project subnetworks, the resulting Betti-number distributions show that the observed values are statistical outliers relative to this simplex-count-preserving, undirected-backbone-preserving null model, rather than automatic byproducts of having many simplices (Unger et al., 2023).

Several related constructions delimit the scope of $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$8:

Construction	Simplices	Distinctive feature
Directed flag complex $$(v_0,\dots,v_k)\in dFl(G) \quad\Longleftrightarrow\quad (v_i,v_j)\in E \text{ for all } i<j.$$9	Ordered transitive directed cliques	Canonically determined by the digraph
Flag tournaplex $v_0$0	All tournaments contained in $v_0$1	Strictly larger than $v_0$2
Path complex	Allowed directed paths	Depends on path closure, not clique closure
Hyperdigraph model	Chosen directed hyperedges	Higher cells need not be forced by the $v_0$3-skeleton

The comparison with path complexes is especially sharp. Path homology and directed-flag-complex homology can disagree even on small digraphs, and the associated Laplacians can distinguish different aspects of directionality. Hyperdigraphs are even more flexible, since higher-dimensional cells can be declared independently of transitive clique conditions. A common misconception is therefore to treat all higher-order topologies on directed graphs as interchangeable; they are not (Jones et al., 2023, Govc, 2020).

A second limitation is that the geometric realization of $v_0$4 can discard directional information that remains visible at the combinatorial level. Directed $v_0$5-connectivity and the resulting preorder spaces were introduced precisely to recover some of that lost asymmetry (Riihimäki, 2022). A third limitation is instability: persistent homology of directed flag complexes is stable under edge subdivision for shortest-path filtrations of weighted DAGs, with interleaving distance bounded by the maximum subdivided edge weight, but non-DAG edge subdivisions and even the addition of a single appendage edge can produce infinite interleaving distance in directed-flag-complex persistence. This contrasts with the more robust behavior of path homology in analogous situations (Chaplin et al., 2024).

Taken together, these results fix the directed flag complex as a precise and nontrivial intermediary between directed graph combinatorics and higher-dimensional topology. Its simplices are simple to define, its homology is computable at scale, its homotopy types can be unexpectedly rich, and its spectral and persistent refinements expose directional phenomena that neither undirected clique complexes nor ordinary persistent homology record fully.