When we talk about trust networks, we usually talk about graphs: nodes and edges, degree distributions, centrality scores. But a graph is a one-dimensional shadow of something richer. The question persistent homology asks is not who trusts whom but what shape does the trust make?
Consider a set of agents in a reputation network. Each agent occupies a position in some abstract space—perhaps determined by behavioral similarity, attestation patterns, or embedding vectors from their interaction history. The distances between agents encode how “far apart” they are in trust space. Two agents who have deeply verified each other sit close together. Strangers sit far apart.
This is a point cloud. And point clouds have topology.
The shape of a genuine trust community is organic: nodes at varying distances, some tight clusters, some loose connections, with the kind of structural heterogeneity that emerges from real social processes. A Sybil cluster—a set of colluding fake identities—has a different shape. It is too regular, too tight, too uniform. Persistent homology can detect this difference by studying how topological features appear and disappear as we vary a scale parameter.
The Vietoris-Rips complex is built by growing a ball of radius ε around each point. When two balls overlap—when the distance between their centers is at most 2ε—we draw an edge. When three points are mutually connected, we fill in a triangle (a 2-simplex). When four points are mutually connected, we have a tetrahedron (a 3-simplex). And so on.
Start with ε = 0. Every point is isolated. As ε grows, edges appear between nearby points. Triangles fill in. Connected components merge. At some point, loops form—cycles of edges that enclose empty space, holes in the simplicial complex. As ε continues to grow, these loops get filled in by triangles, collapsing the holes. Eventually, at large enough ε, everything is one giant simplex.
This sweep through ε is a filtration: a nested sequence of simplicial complexes, each contained in the next. The filtration is the movie. Persistent homology is the summary of what happens in that movie.
In the visualization below, click to add points, or use the buttons to generate networks. Then drag the ε slider and watch the Rips complex grow.
As ε sweeps from 0 to ∞, topological features are born and die. A connected component (an H0 feature) is born when a point first appears and dies when it merges with an older component. A loop (an H1 feature) is born when edges close a cycle without a 2-simplex filling it, and dies when a triangle finally plugs the hole.
The persistence of a feature is the length of its lifetime: death minus birth. Features with high persistence are robust structural properties of the data—they persist across many scales. Features with low persistence are noise: artifacts of particular ε values that vanish almost as soon as they appear.
The persistence diagram plots each feature as a point (birth, death) in the plane. The diagonal line y = x represents zero persistence. Points far above the diagonal are significant. Points near the diagonal are transient. The barcode shows the same information as horizontal bars: long bars are signal, short bars are noise.
Click “Inject Sybil Cluster” above and watch what happens to the persistence diagram.
A Sybil cluster is a set of fake identities controlled by one actor. Because they are generated by a single algorithm, the inter-node distances within the cluster tend to be suspiciously uniform. Real communities have organic distance distributions—some members are close, some are peripheral, distances vary across a wide range. A Sybil cluster is a near-perfect ball of tightly packed points at nearly equal distances.
This uniformity creates a distinctive topological signature. As ε grows and enters the narrow band of Sybil inter-distances, the entire cluster connects simultaneously. Edges, triangles, and higher simplices all appear in a burst. H1 features (loops) are born and immediately killed as triangles fill them in. On the persistence diagram, this manifests as a cluster of H1 points very close to the diagonal—born and dead within a tiny ε-window.
Moreover, the Sybil cluster often creates distinctive H1 features at the boundary where it connects to the genuine network. These loops involve edges between Sybil nodes and honest nodes, and they tend to have a specific persistence range that differs from the organic loops within genuine communities. On the barcode, the Sybil features appear as a stack of short, red bars, visually distinct from the long blue bars of genuine community structure.
Gunnar Carlsson observed that the power of persistent homology comes from studying data not at a single scale, but across all scales simultaneously. The filtration parameter ε is the lens, and by varying it we see structure that no fixed-scale analysis can reveal.
In a Nostr reputation network, the base space is the namespace: different domains of expertise, different relay communities, different interest groups. Trust attestations (kind 30085 events in NIP-XX) carry weights and context tags that place them in this base space. The persistent homology of trust is not a single computation but a family of computations, one for each slice of the base space.
A trust relationship that persists across many domains—an attestation that remains topologically significant whether you filter by “code review,” “content moderation,” or “financial reliability”—is a robust signal. A Sybil cluster that only achieves high persistence in one narrow domain is fragile. The multi-parameter persistent homology of the trust network, viewed as a sheaf over the namespace, gives a far richer picture than any single-domain analysis.
This connects to Carlsson’s broader program of studying data over a base: the fibers (trust structures in each domain) and the sections (trust relationships that cohere across domains) together reveal the full topology of reputation. The H1 features that persist in the product filtration—across both ε and domain—are the features we should take most seriously.
To my knowledge, nobody has deployed multi-parameter persistent homology for decentralized reputation. The single-parameter case demonstrated above is already informative. The multi-parameter case is where the real power lies, and where the mathematics of sheaves and derived categories meets the engineering of trust.