A large strand of social network analysis treats triangles as special. Three people all connected to each other — a closed triad, a triangle in the graph-theoretic sense — show up in the literature as the minimum unit at which social structure is supposed to begin crystallizing. Triangles carry transitivity. They support trust. They coordinate action. They constrain cheating. They are, on the standard account, where the graph becomes more than the sum of its edges. Mark Granovetter’s work on weak ties, Georg Simmel’s treatment of triads, Holland and Leinhardt’s triad census, and the Watts-Strogatz clustering coefficient all assign the triangle a load-bearing role.1 I want to take seriously the question of what is doing the work in that claim.
Three separate things might be doing it. They usually get conflated.
What the triangle carries
The first is visualization. Force-directed layout algorithms — the standard tools for drawing networks, which I have written about in the earlier installments of this series2 — place nodes in two-dimensional space by simulating springs along edges and repulsive forces among nodes. A three-node configuration with all three edges present is, in this physics, the smallest rigid subgraph. You cannot flex it without stretching at least one spring. Larger subgraphs are more floppy: a four-cycle can be squeezed, a path of any length can be straightened, a tree has no rigidity at all. The algorithm’s equilibrium-finding is therefore systematically biased toward arrangements in which triangle-rich regions appear as compact visible clumps. When a network is drawn by a spring simulation, the eye sees clumpiness where there are triangles — not because the triangles are carrying social structure, but because they are carrying physical structure in the visualization. Just take a look at the figure below and compare “how many triangles do you see?” with “how many [insert any other shape with more than 3 corners here] do you see?”
The second is statistical modeling. The standard parametric framework for social networks — the exponential random graph model, or ERGM — conditions on triangle counts, alongside edge counts and star counts, as sufficient statistics for estimating the probability distribution over graphs. The standard descriptive measure of graph-level closure — the clustering coefficient — is a ratio of closed triads to possible closed triads, which is to say, a count of triangles normalized by a count of triangle-candidate configurations. The standard relational-closure notion, transitivity, is operationalized as the proportion of length-two paths whose endpoints are adjacent, which is a count of triangles under a different name. If the models take triangles as sufficient statistics, and the descriptive measures take triangles as numerators, and the cohesion notions take triangles as indicators, then the analyses will return the result that triangles carry structural information. The statistical framework has already decided the answer.
The third is data generation. Social network data most often come from two kinds of sources. The first is name-generator surveys in which egos report their alters, and the alter–alter ties are inferred from ego’s knowledge of what relations her alters have with each other. The second is co-occurrence data in which observed presence in a shared context — a class, a conference, a mailing list, a committee — produces an edge. Both mechanisms over-sample triangles. Ego is more likely to know about an alter–alter tie when ego is directly connected to both alters, so alter–alter triangles are disproportionately reported. Co-occurrence is structurally triangle-producing, because three people who co-occurred pairwise in a bounded context are very likely to have co-occurred jointly in at least one instance, which in the aggregated data typically materializes as the third edge of a triangle. The underlying social process may or may not be triangle-rich; the instrument that samples it is.
These three reinforcing mechanisms produce a finding that is difficult to dislodge. Triangles are rigid in drawings. They are sufficient statistics in models. They are over-sampled in the data. Any literature that combines drawings, models, and survey data — which is to say, all of social network analysis — will return triangle-stability as a robust result regardless of what the underlying social process actually looks like. The finding is not necessarily false. It is, however, unfalsifiable in the weak sense that the apparatus is structurally incapable of returning the finding’s negation even when the truth warrants it. This is the classifier-feedback pattern from Maggie‘s and my AJPS paper3 applied at the level of a research tradition rather than the level of an individual classification decision: the apparatus shapes what counts as a finding, and findings made by the apparatus then feed back into the choices that fix the apparatus in place.
A picture
Before I press further on the apparatus, a picture. The widget below lets you take the smallest non-trivial political networks — named coalitions of three, four, five, seven, eight, and nine people — and try to draw them in two dimensions without edge crossings. Each network is the complete graph on its members: everyone connected to everyone. Drag the nodes; the counter in the corner tracks how many edges cross in your current arrangement. The “Playground” option lets you pick any n up to twelve. The “Prove it” button explains what you’ll find. The “Run the algorithm” button does what it sounds like — tosses the nodes into fresh random positions and lets the force simulation find whatever layout it finds.
Try the Federalist trio first. Then the Quad. Then the P5. Then the Supreme Court. The rest of the post is an argument about what the counter is counting.
The theorem
There is a theorem in elementary graph theory, older than most of the network literature that disregards it. Kuratowski's theorem: a graph is planar — can be drawn without edge crossings in two dimensions — if and only if it does not contain K₅ or K3,3 as a subdivision.4 The complete graph on five vertices, which has ten edges, is the smallest complete graph that cannot be drawn without crossings. K₃, three edges, can. K₄, six edges, can. K₅ cannot.
Euler's formula makes the argument direct. A simple planar graph with V ≥ 3 vertices has at most 3V − 6 edges, because every face in a planar embedding is bounded by at least three edges and every edge borders at most two faces. The complete graph Kn has n(n−1)/2 edges. For n = 4, both quantities equal six — tight, and still planar. For n = 5, the first equals nine and the second equals ten. K₅ has more edges than any planar graph on five vertices can accommodate, and no matter how you arrange the nodes, at least one crossing persists. For n = 9, the deficit is fifteen edges, and a theorem of Richard Guy, confirmed for all n up to twelve, puts the minimum achievable number of crossings at thirty-six — one for each edge in the graph, on average crossed twice. That is the number the widget's counter will not fall below when you try the Supreme Court.
The coalition graph, the planar graph
Consider n actors operating under majority rule. A minimal winning coalition is a subset large enough to prevail on a vote — of size ⌈(n+1)/2⌉ under simple majority, of size q under a q-rule, either way a specific subset-size rather than the whole group. Say that a coalition is internally complete if every two of its members are directly connected — every pair can coordinate without intermediary, every pair is an edge. And ask what graph on n vertices has the property that every minimal winning coalition is internally complete.5
The answer, for any non-unanimity rule, is Kn itself. Because any two vertices lie together in at least one minimal winning coalition, internal completeness of that coalition requires an edge between them, and the condition "every minimal winning coalition is internally complete" therefore forces all n(n−1)/2 edges to be present. For majority rule on three people, that is K₃. On four, K₄. On five, K₅. On nine, K₉. The graph that formalizes the coalition-theoretic property that every minimally empowered subset can coordinate internally is the same graph that Kuratowski's theorem identifies as the boundary of planar representability.
Put the two results next to each other. Up to n = 4, the coalition-theoretic stability graph is also planar — it can be drawn in two dimensions without distortion. From n = 5 onward, the coalition-theoretic stability graph is not planar. The visual intuition that "everyone is connected and the picture is stable" tracks the coalition-theoretic content of that statement exactly up to four actors. At five actors it comes apart. At nine, it comes apart catastrophically: the minimum number of crossings in any drawing of K₉ is thirty-six, which is to say, every single edge of the graph is crossed by other edges somewhere in every possible drawing, on average twice. The apparatus does not report this. The apparatus reports triangles.
Where the apparatus converges, and where it does not
The closure tradition's use of the triangle as the unit of analysis is not arbitrary and it is not obviously wrong. It is privileging the case in which the visual apparatus, the statistical apparatus, the data-generation apparatus, and the coalition-theoretic substance happen to coincide. A closed triad is rigid in force-directed layouts. It is a sufficient statistic in ERGM. It is over-sampled by ego-alter surveys. And it is the minimal-winning-coalition graph of the smallest interesting case of majority rule. All four of these things are true of triangles. None of them is true of pentagons, or K₅, or K₉, or any of the structures that formalize coalition stability for bodies larger than four. The apparatus of social network analysis has gravitated to the one structure where its priors and its targets happen to agree.
This is why the P5 is the case where the critique is sharpest. The United Nations Security Council's five permanent members are almost exclusively analyzed in coalition-stability terms — veto overlap, bloc alignment, abstentions-as-defections, the arithmetic of which subsets can and cannot block resolutions. The P5 has five members. K₅ is the smallest non-planar complete graph. Any network diagram of P5 interactions ever drawn has at least one edge crossing in it, because every drawing of K₅ must. That crossing is not a feature of the five nations' actual relationships. It is a feature of the two-dimensional space the relationships are being projected into.
And this is why the Supreme Court is the case that does the most damage to standard commentary. Political analysis of the Court routinely leans on triangle-coalition language — the 3-3-3 Court, the conservative bloc, Roberts as the fulcrum, triangles around a median justice in earlier eras — as if the elementary unit of analysis for a nine-member body were the triadic sub-coalition. The Court has eighty-four possible minimum winning coalitions of five justices, and the condition that every one of them be internally complete produces exactly K₉, which has a minimum of thirty-six crossings in any two-dimensional drawing. This does not mean the Court lacks structure. It means the structure the Court has cannot be read off a picture of it, and the triangle language routinely used to describe it is inherited more from the apparatus that draws the picture than from the institution the picture is of.
What the series is doing
Conservation of impossibility shows up here in the form it has shown up elsewhere on this blog. The representational bias is not escaped by moving to richer formalism. Replace the graph with a simplicial complex — a generalization in which triangles, tetrahedra, and higher-dimensional simplices are all elementary objects — and the question of which simplices are "real" is relocated into the choice of which dimension to include. Replace the graph with a hypergraph in which edges can connect arbitrary subsets of vertices, and the question of which hyperedges to include is relocated into the data-generation assumption about which subsets are observable. The apparatus has a prior. The choice of apparatus is therefore a choice of prior, and the prior comes along even when you pick the fancier apparatus, because you had to decide what the elementary objects of the fancier apparatus were going to be, and that decision is not free.
What this series is building, post by post, is not an argument to abandon network analysis. Maggie and I use it; political science needs it; every one of the recent methodological fights that matter in our field turn on networks somewhere. The series is building a catalog of the things the apparatus cannot see. The first post was about what the drawing cannot see. The second was about what the narrative cannot see. This post is about what the triangle cannot see. Each installment finds the same structural pattern in a different domain, and the pattern is this: the representation has a prior, the prior rewards certain findings, and the findings feed back into the data that validate the representation. Classifiers shape the populations they classify. Drawings shape the intuitions they illustrate. Units of analysis shape the phenomena that fit within them. The work is not to escape the apparatus. The work is to know where the apparatus is looking and where it is not.
With that, I leave you with this.
1 Granovetter, Mark. "The Strength of Weak Ties." American Journal of Sociology 78(6): 1360–1380, 1973. Simmel's discussion of triads appears across his writing on the sociology of group size; the canonical English-language gathering is Kurt Wolff's translated volume The Sociology of Georg Simmel, Free Press, 1950. Holland, Paul W. and Samuel Leinhardt. "A Method for Detecting Structure in Sociometric Data." American Journal of Sociology 70(5): 492–513, 1970. Watts, Duncan J. and Steven H. Strogatz. "Collective Dynamics of 'Small-World' Networks." Nature 393: 440–442, 1998. A careful reader will note that Ronald Burt's work on structural holes is closure-tradition-adjacent but argues in the opposite direction from the sources above — that absence of triangles (brokerage positions) is the structural feature worth analyzing. The critique here is of the closure tradition as a tradition, not of Burt's specific argument, though both share the feature that triangles are what the apparatus has made legible.
2 The Physics of Political Networks, April 1, 2026, and The Dangers of Graphic Expression, April 20, 2026 (Ed.: Oh, I see what you did there, DOGE, I'm a little slow...). The first defined NEIIA and the three-tier robustness hierarchy over node positions, betweenness centrality, and degree centrality. The second extended the critique from algorithmic layouts to narrative layouts using the DOGE case.
3 Penn, Elizabeth Maggie and John W. Patty. "Classification Algorithms and Social Outcomes." American Journal of Political Science, 2025. The paper's central result is that accuracy-maximizing classifiers can exhibit negative threshold rules under endogenous base rates — classifying positive precisely where evidence is weakest — because classifiers do not only measure populations, they reshape them through the incentives classification creates. The feedback structure described there is individual-scale and decision-theoretic; the analogy to a research apparatus is not identical to the formal result but preserves its essential structure.
4 Kuratowski, Kazimierz. "Sur le problème des courbes gauches en topologie." Fundamenta Mathematicae 15: 271–283, 1930. The crossing-number result used in the text is due to Guy (1972), posed as a conjecture and confirmed by computer-assisted proof for all n ≤ 12 in work culminating in McQuillan, Pan, and Richter, "On the crossing number of Km,n and Kn," Journal of Combinatorial Theory, Series B, 2015. The formula is cr(Kn) = ¼⌊n/2⌋⌊(n−1)/2⌋⌊(n−2)/2⌋⌊(n−3)/2⌋. For n = 9, this evaluates to 36. Whether the conjecture holds for all n is open.
5 The claim that the coalition-stability graph is Kn presumes not just anonymity but self-duality — the property that a coalition wins if and only if its complement loses. This is a defining property of simple majority rule by May's theorem, and it is what forces every pair of voters to appear together in some minimal winning coalition. Under supermajority rules, under weighted voting, or under any rule that fails self-duality, the minimal decisive coalitions and the minimal blocking coalitions come apart, and the required graph can be strictly sparser than Kn. Which classes of non-self-dual rules admit planar coalition-stability graphs is, to my knowledge, an open question, and one Maggie and I have begun noodling with. May, Kenneth O. "A Set of Independent Necessary and Sufficient Conditions for Simple Majority Decision." Econometrica 20(4): 680–684, 1952. For a more recent treatment of why majority rule is the rule it is, see Dasgupta, Partha and Eric Maskin, "On the Robustness of Majority Rule," Journal of the European Economic Association 6(5): 949–973, 2008.