Simplicial Embeddings in Topology & Machine Learning
- Simplicial embeddings are maps that preserve the incidence and homological structure of simplicial complexes when embedded in geometric spaces.
- They enable various methodologies including isometric, spectral, and simplex-constrained latent embeddings in fields like topology, geometry, and machine learning.
- Applications range from trajectory analysis and reinforcement learning to complex representation learning, addressing challenges such as high-dimensional structure preservation and computational complexity.
Searching arXiv for relevant papers on “simplicial embeddings” across topology, geometry, and machine learning. Simplicial embeddings denote a family of constructions in which simplices, simplicial complexes, higher-order signals, or simplex-constrained latent codes are placed into another geometric or representation space while preserving selected structure. In the cited literature, the term covers continuous, PL, linear, geodesic, and isometric embeddings of simplicial complexes into , , and Minkowski space, as well as spectral embeddings of edge-flows, Hasse-diagram embeddings of simplices, and latent representations constrained to products of probability simplices (Frick et al., 2021, Minemyer, 2012, Frantzen et al., 2021, Lavoie et al., 2022).
1. Formal scope and recurring mathematical structures
A simplex is the convex hull of finitely many affinely independent points, and an abstract simplicial complex is a family of simplices closed under taking faces. In combinatorial treatments, $0$-simplices are vertices, $1$-simplices are edges, $2$-simplices are triangles, and higher-dimensional simplices encode higher-order interactions. Several of the cited works use the canonical probability simplex
either as a geometric object in its own right or as a constraint set for latent variables (Billings et al., 2019, Ren et al., 2023, Obando-Ceron et al., 15 Oct 2025).
A second recurrent structure is the simplicial incidence calculus. For a simplicial complex with vertex–edge incidence matrix and edge–triangle incidence matrix , the unnormalized Hodge $1$-Laplacian is
This operator governs edge-flows and yields the Hodge decomposition
0
where the harmonic space 1 is isomorphic to the first homology and therefore records holes in the underlying complex (Frantzen et al., 2021).
A third recurring meaning is architectural rather than topological. In simplex-constrained latent models, an encoder output is reshaped into 2 blocks of size 3, and each block is mapped by a temperature-scaled softmax to 4: 5 The resulting representation lies in 6, and each block can be interpreted as barycentric coordinates over a learned set of vertices (Obando-Ceron et al., 15 Oct 2025, Lavoie et al., 2022).
2. Embeddings of simplicial complexes into geometric spaces
In topological and metric geometry, simplicial embeddings are maps from the geometric realization of a simplicial complex into a sphere, Euclidean space, or pseudo-Euclidean space that preserve the simplicial structure in an appropriate sense. The strongest forms are simplicial isometric embeddings, where the edge-squared lengths prescribed on the complex agree with the pullback of the ambient bilinear form. For an indefinite metric polyhedron, a simplicial isometric map into 7 preserves the signed squared lengths
8
where 9 is the Minkowski inner product (Minemyer, 2012, Galashin et al., 2015).
A sharp few-vertex sphere-embeddability criterion is available for complexes on $0$0 vertices. If $0$1 is a simplicial complex with $0$2 and $0$3 denotes its family of minimal non-faces, then
$0$4
For this regime, continuous embeddability into $0$5 or $0$6 is equivalent to geodesic or linear embeddability, yielding a Fáry-type rigidity phenomenon for complexes on $0$7 vertices (Frick et al., 2021).
For indefinite metric polyhedra, existence theorems depend on the maximal vertex degree or degeneracy. A compact $0$8-dimensional indefinite metric polyhedron with maximal degree $0$9 admits a simplicial isometric embedding into $1$0, where $1$1, and a PL isometric embedding into $1$2 when simpliciality is relaxed (Minemyer, 2012). A later constructive result shows that a possibly non-compact $1$3-dimensional $1$4-degenerate indefinite metric polyhedron admits a simplicial isometric map into $1$5, and if $1$6 then a simplicial isometric embedding into $1$7. The same work proves an extension theorem: every partial simplicial isometric embedding in $1$8-general position extends to a simplicial isometric embedding of the whole space (Galashin et al., 2015).
A different constructive direction is provided by parsimonious cones. For a finite $1$9-dimensional simplicial complex $2$0 with $2$1 simplices and any $2$2, there exists a triple $2$3 of simplicial complexes with bounded degrees such that $2$4, $2$5 is homotopy equivalent to $2$6, $2$7, and $2$8 and $2$9 are bounded by 0. A related local simplification theorem produces a homotopy equivalent bounded-degree complex 1 with 2 (Berdnikov, 2022).
3. Local criteria, rigidity, and extremal consequences
For 3-complexes in 4-space, embeddability can be characterized through local rotation data. A finite simply connected 5-dimensional simplicial complex embeds in 6 if and only if it has a planar rotation system, meaning that every vertex link graph embeds in the plane with edge-compatible cyclic orders. Under the additional condition 7 for some prime 8, embeddability is equivalent to the conjunction of 9 and existence of a planar rotation system (Carmesin, 2017).
Codimension-one embedding imposes a distinct homological obstruction. If a 0-dimensional simplicial complex 1 PL-embeds in 2, then 3 has a 4-complete basis. Writing 5 for the girth, defined as the minimum number of 6-faces in a subcomplex with nontrivial top homology over 7, the paper derives the inequality
8
Combined with Euler–Poincaré and strong Morse inequalities, this yields upper bounds on 9 in terms of lower-dimensional face numbers and Betti numbers (Björner et al., 2016).
For complexes embeddable in the Euclidean 0-space, links of vertices satisfy a stronger higher-order restriction. If 1 is a PL embedding of a 2-complex, then for any three vertices the intersection of their link-complexes is linklessly embeddable in 3. This leads to the upper bound
4
for the maximum number of 5-simplices in an embeddable 6-complex on 7 vertices. In the case 8, the triple intersection of link-complexes is in fact planar (Parsa, 2015).
In geometric group theory and low-dimensional topology, “simplicial embedding” often means an injective adjacency-preserving map between combinatorial complexes associated to a surface. For multiarc graphs 9, if $1$0 are non-exceptional, $1$1, $1$2, and $1$3, then any simplicial embedding $1$4 is induced by a $1$5-injective embedding $1$6 together with a fixed $1$7-multiarc, and its image is strongly convex (Parlier et al., 2019). An analogous theorem holds for $1$8-multicurve graphs $1$9, where simplicial embeddings are induced by 0-injective embeddings of surfaces and a fixed multicurve, and the automorphism groups satisfy 1 under the stated complexity hypotheses (Erlandsson et al., 2015).
4. Spectral, Hodge, and neural representation embeddings
A spectral interpretation of simplicial embeddings appears in trajectory analysis on discretized manifolds. A trajectory is encoded as an edge-flow 2, where each oriented edge receives 3, 4, or 5 according to traversal direction. If 6 contains the harmonic eigenvectors of the Hodge 7-Laplacian 8 corresponding to zero eigenvalues, then the flattened simplicial embedding is
9
Each coordinate records circulation around a homological cycle, so the embedding is low-dimensional and topologically interpretable. In the synthetic experiments reported in the paper, Local Outlier Factor and Isolation Forest identify planted outliers in this harmonic space, and in the whale dataset the method isolates trajectories with atypical circulation around Hudson Bay and Hudson Strait (Frantzen et al., 2021).
Simplex2Vec adopts a different route. It performs random walks on the Hasse diagram of a simplicial complex and trains a skip-gram model with negative sampling on the resulting sequences, thereby embedding vertices and higher-dimensional simplices jointly. The method uses either uniform transitions or bias schemes that favor higher-order or lower-order simplices, and it was evaluated on Costa–Farber random complexes, sociopatterns school data, and fMRI connectivity. The reported results show that increasing maximal simplex dimension can materially change persistent-homology summaries and improve community detection when higher-order bias is used (Billings et al., 2019).
Entire-complex representation learning extends the same idea from simplex-level to complex-level embeddings. In “Simplicial Complex Representation Learning,” geometric message passing uses adjacency, coadjacency, boundary, and coboundary operators to produce simplex embeddings 00, and then pools them into a complex embedding
01
Training minimizes
02
so that Euclidean distances in the learned embedding space approximate a chosen complex-to-complex distance 03. On the deformation-transfer mesh dataset, the learned 04-dimensional embeddings cluster shapes by category and generalize to unseen categories (Hajij et al., 2021).
The same Hodge-theoretic perspective has also been used for semantic-native communication. In that setting, a transmitter maps data to a 05-order simplicial complex, computes simplicial embeddings through a simplicial convolutional autoencoder, and transmits the latent representation through an AWGN channel. The receiver reconstructs both the simplicial Laplacians and the simplex features. On a Semantic Scholar Open Research Corpus experiment, the reported system reconstructs missing data with 06 accuracy, remains stable under channel noise, and outperforms a conventional linear autoencoder that does not use simplicial structure (Zhao et al., 2022).
5. Simplex-constrained latent embeddings in machine learning
A recent line of work uses “simplicial embeddings” to denote latent vectors explicitly constrained to products of probability simplices. In self-supervised learning, a representation 07 is split into 08 blocks of dimension 09, each block is mapped by temperature-scaled softmax to 10, and the final representation 11 lies in 12. The cited theory shows that for downstream classification the SEM-normalized representation yields a tighter generalization bound than an unnormalized baseline, with the complexity term improving as 13. Empirically, SEM improves linear-probe performance for SimCLR, MoCo-v2, BYOL, Barlow Twins, SwAV, DINO, and VicReg on CIFAR-100, and improves BYOL on ImageNet from 14 to 15 top-1 while also improving ImageNet-V2, ImageNet-R, ImageNet-C, and ImageNet-A robustness benchmarks (Lavoie et al., 2022).
The same product-of-simplexes construction has been combined with iterated learning for compositional generalization. Here a teacher produces one-hot pseudo-labels sampled from each simplex block, a student matches them by a multi-label cross-entropy imitation loss, and then the student is fine-tuned on the downstream task before becoming the next teacher. The motivation is explicitly tied to Kolmogorov complexity: compositional bijections admit shorter descriptions than arbitrary bijections, and iterated learning biases the model toward compressible, reusable codes. The reported results show that neither SEM-only nor iterated-learning-only suffices on the controlled vision tasks, whereas the combination improves compositional generalization and also improves molecular property prediction on ogbg-molhiv, ogbg-molpcba, and PCQM4Mv2 (Ren et al., 2023).
In actor–critic reinforcement learning, simplicial embeddings are used as lightweight representation layers placed at the penultimate layer of an actor or critic network. Each block is constrained to a simplex by softmax, producing sparse, bounded, and often near-discrete features. The paper argues that this geometric inductive bias preserves feature covariance trace, stabilizes effective rank under drift, reduces neuron dormancy, and keeps gradient energy large. When inserted into FastTD3, FastSAC, and PPO, SEM improves sample efficiency and asymptotic return across HumanoidBench, IsaacGym, MTBench MT50, Atari-10, and a Booster T1 sim-to-real setup, with effectively no runtime penalty. The strongest gains are reported for actor-only placement at the penultimate layer, and the method uses standard softmax rather than straight-through estimators, sparsemax, or entmax (Obando-Ceron et al., 15 Oct 2025).
6. Applications, limitations, and open directions
Combinatorial surface theory supplies another application domain. A simplicial surface can be encoded through incidences of vertices, edges, and triangular faces, and its face graph is cubic. In this setting, umbrella cycles define a cycle double cover, and embeddings of cubic graphs on simplicial surfaces correspond to strong graph embeddings. For 16-connected cubic planar graphs, the induced embedding on a vertex-faithful simplicial sphere is unique up to isomorphism; the same paper classifies embeddings on simplicial projective planes, tori, and Klein bottles by 17-orbits of specified subgraphs in the planar dual 18 (Weiß et al., 2024).
The literature also records sharp negative results. At the geometric level, the complexity of embeddings can grow exponentially even when embeddability is topologically unobstructed. For each 19, there are families of 20-complexes 21 with bounded local combinatorial complexity such that every embedding has thickness at most 22, and the refinement complexity satisfies 23. In codimensions 24 and 25, and in a relative codimension-26 setting, the paper further shows non-recursive lower-bound phenomena for refinement complexity (Freedman et al., 2013).
At the algorithmic and modeling level, the limitations are domain-specific. Spectral and Hasse-diagram methods depend on the correctness of the underlying simplicial complex, so coarse discretization, mis-specified holes, or high maximal simplex order can blur the intended structure or make Hasse construction expensive (Frantzen et al., 2021, Billings et al., 2019). Simplex-constrained latent models introduce hyperparameters such as the number of blocks, block size, and temperature, and the cited reinforcement-learning work notes that benefits may be smaller under extreme distribution shift or very sparse rewards, while the self-supervised and iterated-learning papers emphasize sensitivity to 27, 28, and 29 (Obando-Ceron et al., 15 Oct 2025, Lavoie et al., 2022, Ren et al., 2023).
These uses suggest that “simplicial embeddings” is not a single method but a structural principle. In topology, the principle is preservation of incidence, homology, and ambient nonintersection. In spectral and representation-learning settings, it is preservation of higher-order adjacency, harmonic circulation, or simplex-constrained barycentric structure. What remains common across the cited literature is that simplices are treated as first-class objects rather than as reducible collections of pairwise edges, and the embedding is designed to retain that higher-order organization in the target space.