Read the graph the other way and each node is a polyhedron, each link a shared face — space as a packing of grains
Everything so far has treated a spin network as a graph: nodes, links, labels, combinatorics. That reading is complete and correct, and it is also almost impossible to think with. This lesson gives you the other one. Read the same state from the other side and each node becomes a grain of space, each link becomes a face that two grains share, and the spin on that link is how big the shared face is.
The bridge between the two readings is a theorem from 1897, proved by Minkowski for reasons that had nothing to do with physics. Give him a collection of vectors whose lengths are the areas you want your faces to have, and whose sum is zero, and he hands back exactly one convex polyhedron with those outward face-area vectors. Not one of many. Exactly one.
Now look at the condition he needs. The vectors have to sum to zero. That is the Gauss constraint, the demand that the node be invariant under rotations of the internal frame, which we imposed two lessons ago for reasons of pure representation theory. A gauge-invariant node and a grain of space turn out to be the same object satisfying the same equation, and nobody arranged that. It is the single most persuasive coincidence in the subject.
Four spins fix four face areas, and the closure condition does the rest. Step through the allowed recoupling labels and you are stepping through an orthonormal basis of the node’s Hilbert space, one basis vector at a time. Drag the angle and you are sliding along the classical shape space that basis quantises. Watch the closure defect stay at zero while the solid changes shape underneath it.
The closure defect stays at machine zero however you move the shape: closure is built into the construction, which is the point. Stepping k walks through an orthonormal basis of the intertwiner space, and the number of steps is exactly its dimension, so the recoupling label you met as an algebraic bookkeeping device is the area of a real internal face. Sweeping φ moves along the other coordinate of the classical shape space, keeping all four face areas fixed while the volume and the dihedral angles change. Drag to orbit.
Count the freedom in a tetrahedron with four fixed face areas. Four vectors in three dimensions is twelve numbers; fixing the four lengths removes four; imposing closure removes three; quotienting by the rotations that carry one closing configuration to another removes three more. Two are left. The shape space of a quantum tetrahedron is two-dimensional, which means it is a phase space, which means it can be quantised in the ordinary way and will produce a finite-dimensional Hilbert space.
Barbieri did exactly that in 1998, and the answer was the intertwiner space we already had. The correspondence is tighter than a rough analogy: the two coordinates on the classical shape space match the two labels on the quantum side. One of them is the length of the sum of the first two area vectors, geometrically the area of an internal face slicing the tetrahedron between the (1,2) pair and the (3,4) pair. That is precisely the recoupling label k of the four-valent intertwiner, and its allowed range is the same triangle-inequality range in both descriptions. The other coordinate is a dihedral angle, and it is canonically conjugate to k. The classical shape space and the quantum basis are not similar. They are the same space, counted twice.
Volume follows from the shape. Once the area vectors close, the tetrahedron they determine has a definite volume, and the formula for it in terms of those vectors is exact and short. This is the semiclassical face of the volume operator from the previous lesson: the operator’s large-spin eigenvalues track the volume of the polyhedron whose faces carry the corresponding areas, even though the operator itself is a matrix on the intertwiner space rather than a formula in the areas.
Turn the four area vectors freely and watch the residual. While it is nonzero there is no grain of space, only four arrows that fail to meet. Close them and a tetrahedron appears, uniquely. Then rotate all four together and notice that nothing measurable moves: the node carries a shape but never an orientation, because there is no surrounding space for it to be oriented in.
Watch the red residual, not the arrows. While it is nonzero the four faces refuse to shut and the right-hand pane stays empty; the moment it vanishes a tetrahedron appears, and by Minkowski’s theorem it is the only one with these four face-area vectors. Then drag the last slider: rotating all four vectors together leaves the defect and the volume untouched. That leftover global SU(2) is exactly what the Gauss constraint quotients out, which is why a node carries a shape and no orientation. Drag to orbit.
It is worth being careful here, because this is the point where the subject is most often oversold. The polyhedra are not little solids sitting inside a container. There is no container; that was the whole argument of the second lesson. They are not small in the sense of being tiny objects placed at tiny coordinates, because they have no coordinates. And a spin network state is not a picture of a crystal.
What is actually true is narrower and still remarkable. The intertwiner at a node has the same phase space as the shape of a polyhedron with those face areas, so everything the node can be is something a polyhedron can be. The adjacency of the graph says which grains touch. And when a graph is large and its spins are large, this is how you reconstruct a semiclassical geometry from the state: you build the grain at each node and glue them along shared faces. The picture earns its keep as a reconstruction procedure, not as a claim about what space looks like under a strong enough microscope.
One consequence is pleasant. Volume lives at the nodes and area lives at the links, and in the polyhedral reading that is not two separate facts but one: a grain has a volume, and the faces where grains meet have areas. A three-valent node has three faces, which cannot close into a solid, which is exactly why the volume operator annihilates it.
The same small network drawn both ways, with a slider to move between them. Change a link’s spin and the face it stands for grows in both of the grains that share it. Then turn on the twist display and look closely at a shared face: the two neighbours agree about its area and disagree about its shape. That mismatch is not a drawing artefact, and it is the honest caveat of this lesson.
Slide left for the graph, right for the grains: five four-valent nodes, ten links, and every link a face shared by the two polyhedra it joins. Raise a link’s spin and the highlighted face grows in both neighbours at once, because the same spin is the area of the same face read from either side. Then switch on the twisted geometry check. The two grains agree on the face’s area exactly and disagree about its shape, since each builds it from its own four spins and its own intertwiner. That is why a generic spin network is not a Regge geometry of flat blocks glued face to face, and it is an open problem, not a settled one. The grains are drawn at their nodes with faces turned towards their neighbours as nearly as one rotation allows; the layout is a drawing aid, since a spin network has no positions of its own.
Here is the problem the last demo is showing you. Each node reconstructs its own grain from its own four spins and its own intertwiner. Two neighbouring nodes share a link, so they agree about the area of the face between them, because that area is the link’s spin and there is only one link. They do not in general agree about anything else. The face has one shape as computed by the grain on the left and a different shape as computed by the grain on the right, and there is no rule in the kinematical theory forcing the two to match.
A geometry built from flat pieces glued cleanly edge to edge is a Regge geometry, and Regge geometries are the natural discrete stand-in for a smooth spacetime. What a generic spin network describes is not one of those. Freidel and Speziale named it in 2010: a twisted geometry, in which the gluing matches areas but not shapes. Regge geometries sit inside the space of twisted geometries as a measure-zero subset.
Whether this is a defect or a discovery is genuinely open. One reading is that the extra freedom is physical, and that quantum geometry is simply more general than the piecewise-flat picture we imported from classical discretisations. Another is that the semiclassical limit has to suppress the twisting somehow, and that until someone shows how, the reconstruction of smooth space from these states is incomplete. This is not a footnote. It is one of the specific technical reasons the semiclassical limit remains the programme’s hardest open problem, and we will return to it in the final lesson.