Twenty-four explorations have passed through this representation. Some patterns emerge that weren't obvious at the start.
The series began with Inselberg's point–polyline / line–point duality from the 1980s. Articles 1–4 covered the foundations: crossings, hyperplane signatures, axis permutations. Articles 5–7 moved outward to bundles, containment, and brushing. Articles 8–11 covered applied domains.
Articles 12–14 went into 4D: regular polytopes, 4D invariants, and limits at high dimension. Article 15 covered exceptional structures (the math lives in the Exceptional Atlas and Modular Forms). The tail covered quasicrystals, permutohedron and associahedron, Cayley graphs, knot invariants, error-correcting codes, and the Birkhoff polytope.
This article collects observations that emerged along the way — patterns that aren't obvious from any single article but appear repeatedly across them.
Brushing constrains an axis to a sub-interval. The cut-and-project method for quasicrystals selects lattice points whose perpendicular coordinates fall inside an acceptance window. These are the same operation.
In the cut-and-project view, you split the coordinates of an N-dimensional lattice into a "physical" part (the coordinates you will project onto) and a "perpendicular" part (the coordinates you will constrain). A lattice point is admitted if its perpendicular coordinates land inside a bounded window. If the projection direction is irrational relative to the lattice basis, the projected points form a quasicrystal. What looks aperiodic in the physical space is not aperiodic at all; it is a range constraint in the full space.
This is, precisely, axis brushing. Take a lattice and draw its points in 2-axis parallel coordinates. Brush the perpendicular axis to a narrow window. The surviving polylines correspond to lattice points whose perpendicular coordinate lies in the brushed range. Project their physical coordinate onto a line. You have just computed a Fibonacci chain by brushing.
Figure 1. Left: a tilted lattice in 2-axis parallel coordinates. Lines connect each lattice point's parallel and perpendicular coordinates. Drag the window to brush the perpendicular axis. Right: the parallel coordinates of the admitted points, projected onto the physical direction, produce the Fibonacci chain: long (L) and short (S) intervals in the golden-ratio proportion. Toggle the framing labels: the operation is the same either way.
Cut-and-project calls the window an "acceptance region"; brushing calls it a "selected range." The terminology differs; the operation doesn't.
The indexed points between adjacent axes (intersections of polyline segments, from article 1) characterize a line in N-space completely. For polytopes, indexed points cluster into the polytope's combinatorial structure.
For any structured point set, the indexed-point distribution forms a distinctive texture. Random data: a cloud. The tesseract: a cross-pattern from ±1 coordinates. E8 root system: dense concentric arcs from constant root length √2 and limited pairwise angles. Hamming codewords: a lattice of isolated points from binary coordinates and linear constraints.
Each texture is a fingerprint of the underlying structure type.
Figure 2. Indexed-point signatures for four structures. Each panel shows the polylines in pale blue and every indexed point (intersection of two polylines between an adjacent pair of axes) as a dot. Random data gives a diffuse cloud. The tesseract gives a sparse orthogonal pattern. The E8 roots give dense layered arcs. The Hamming code gives a rigid grid with rectangular symmetry. Same operation, four different textures.
The indexed-point locations are coordinates of the dual object in a derived space. A structured set of lines has a correspondingly structured set of dual points.
Recall the convergence-point formula from article 1:
Where the convergence point falls tells you the type of relationship between two variables. If , convergence races off to infinity and the two variables are strongly positively correlated. If , convergence sits at the midpoint between the axes and the variables are negatively correlated. The location, not the magnitude, encodes the relationship.
Apply this to a root system. For each pair of adjacent axes in a parallel coordinates display, and for each pair of roots, compute the slope between the values on those axes, and plug it into the formula. The collection of resulting convergence points, one for each pair of roots and each adjacent axis, is a scatter. The scatter is different for each root system type. For the classical families , , the convergence points cluster along a few distinctive lines. For the exceptional F4, E6, E7, E8, the distribution is richer, because the pairwise slopes take more distinct values.
Figure 3. Convergence-point scatter for six root systems. Each dot is a convergence point for a pair of roots between an adjacent pair of axes. The distribution forms a signature: produces a few tight clusters; the exceptional systems produce layered patterns with more distinct slope values. The x-axis is horizontal position (slope-dependent), the y-axis is vertical position (intercept-dependent). Toggle between systems to compare.
The convergence formula was a tool for recognizing two-variable correlation. Applied to root systems, it classifies type — An, Dn, or exceptional — by scatter signature.
The 240 roots of E8 are the kissing vectors of the densest 8-dimensional sphere packing. The 196,560 minimal vectors of the Leech lattice are the kissing vectors of the densest 24-dimensional packing. The packing densities were proved optimal in 2016: Viazovska for E8, then Cohn, Kumar, Miller, Radchenko, and Viazovska for Leech. The kissing numbers themselves were already known from the 1979 Odlyzko-Sloane and Levenshtein bounds.
In parallel coordinates, this optimality shows up as a density pattern. The coordinate values at each axis are not uniformly distributed; they concentrate on a finite set. For E8, the coordinates live in ; the marginal distribution on any axis is a handful of spikes. For the Leech lattice (scaled to have integer entries), the coordinates live in with particular frequencies imposed by the underlying Golay code structure.
The density of polylines passing through a value on an axis reflects how often that value appears among the kissing vectors. And the joint density between adjacent axes reflects the constraint structure of the lattice. Sphere packing is a density condition, and parallel coordinates displays density directly.
Figure 4. Left: all 240 roots of E8 as polylines on 8 axes, with the marginal density at each axis shown as a small histogram. Right: 500 randomly sampled minimal vectors of the Leech lattice on 24 axes. Toggle the density overlay. The density spikes reflect the discrete coordinate values enforced by the kissing constraint. E8 concentrates on 5 values per axis; the Leech sample spreads across more, with the underlying Golay code structure shaping the marginals.
The algebraic structure of a group appears in parallel coordinates as visual bundles. S4's 24 elements split into five conjugacy classes (identity, transpositions, double transpositions, 3-cycles, 4-cycles). S5's 120 elements split into seven bundles, one per partition of 5. Each bundle has the predicted class size and a characteristic pattern.
The same plays out in root systems: E8's 240 roots split into Weyl-group orbits, with integer- and half-integer-coordinate roots forming distinct bundles. For monstrous moonshine, McKay-Thompson coefficients bundle by Monster conjugacy class.
Conjugate elements are interchangeable under inner automorphism, and the interchangeability translates to polyline similarity.
Figure 5. Permutations of 4 elements (left) and 5 elements (right), drawn as polylines with one axis per position. Colors mark conjugacy class: identity, transpositions, double transpositions, 3-cycles, 4-cycles, and their additions (5-cycles, and compositions of a transposition with a 3-cycle). Each bundle has the predicted conjugacy-class size. The group structure is the cluster structure.
Inselberg's point–line duality is one example of a broader pattern. Every area this series touched has its own duality, and each leaves a visible relationship in parallel coordinates.
The parallel coordinates representation of each duality looks like the representation of every other: a reflection, permutation, or inversion of the polyline structure. Duality becomes a recognizable style of operation.
Figure 6. Four dualities, four visible relationships. Tesseract and 16-cell: polylines exchanged via a coordinate swap. Hamming code and its dual: the information and parity blocks swap roles. Trefoil and mirror trefoil: invariant profiles differ in a single sign. A root system (drawn for D4) and its dual: lengths and co-lengths exchanged. Each duality is a specific kind of reflection.
The parallel coordinates transformation maps geometric objects (points, lines, flats, polytopes) to other geometric objects (polylines, indexed points, polyline families, dual patterns), preserving some structure and breaking other.
Preserves: incidence (incident flats produce intersecting polylines at predicted indexed points), dimension, and the "same line" relation. Breaks: rotational continuity (rotation in the source becomes nonlocal in the target) and translation locality.
Projective transformations on the source correspond under duality to projective transformations on the target. Parallel coordinates intertwines them — a self-functor on the category of projective flats.
This suggests cataloging other visualization primitives that behave functorially: Radon transform, stereographic projection, and parallel coordinates each give a different dual.
Not every structure fit. The later articles kept running into walls. Each names a limit of the representation and a direction for further work.
Density. The Leech lattice has 196,560 minimal vectors. You cannot draw them as polylines. We sampled 500 and showed the rest as statistical aggregates. For objects of this density, parallel coordinates reduces to a density texture, essentially a picture of the marginal distributions. Most of the combinatorial structure is invisible. We need a different representation for dense discrete sets.
Size of the data matrix. Monstrous moonshine involves a character table. Parallel coordinates with 194 axes is unreadable; beyond approximately 12 axes, individual polylines cease to be distinguishable. We showed fragments: low-order Fourier coefficients, selected conjugacy classes. The full object was always behind a wall.
High dimension. Even E8, at 8 dimensions, is near the limit of what the representation can show clearly. Beyond dimension 12 or so, polylines crowd together and the eye cannot follow individual lines. The curse of dimensionality is not specific to parallel coordinates (it is a general perceptual limit), but parallel coordinates reveals it early.
Complex values. Representation theory and modular-form theory use complex numbers as first-class citizens. A character value is a complex number; a Fourier coefficient is a complex number; a Hermitian inner product takes complex values. Parallel coordinates natively shows real values. Showing complex values requires a choice: modulus and phase as two separate axes, real and imaginary as two separate axes, or something more creative. None of the choices are canonical, and all of them double the axis count.
Continuous objects. Smooth manifolds, Lie groups as topological spaces, fundamental groups and their covering spaces: these are continuous or topological objects, not sets of discrete points with numerical coordinates. We can sample them, but the sample is not the object. For continuous structures, the representation is a lossy approximation.
The series isn't complete. Each limitation above suggests a direction.
Scaling to hundreds of axes. The representation needs graceful degradation: hierarchical grouping (axes in named sub-blocks), attention-based highlighting (3–4 most relevant axes per selected polyline), or coupled mini-maps (low-res overview + high-res detail).
Parallel complex coordinates. Each axis a complex value; polylines become chains of 2D points. The duality needs reworking. Possible application: character tables and modular forms.
Group-equivariant parallel coordinates. If the underlying data has a known symmetry group, the visualization should respect it. An equivariant layout would arrange axes so that the group action becomes a visible cyclic, reflective, or swap symmetry in the polyline pattern. For root systems, this would mean arranging the axes in the order of a Coxeter element's action, so that the rotation of the Weyl group becomes a visible rotation in parallel coordinates.
Topological parallel coordinates. Move beyond points in to chains and cycles of a simplicial complex. A polyline would represent not a point but a path in the complex. Intersections would represent boundary relations. This would push parallel coordinates toward homology, and might give a visual language for persistent homology in applied topology.
Inverse problems. Given an unfamiliar polyline pattern, can we reconstruct the mathematical structure that produced it? The connections in this article suggest that such reconstruction is feasible for well-known structures; the textures of the tesseract and E8 and the Hamming code are distinguishable. Could we train a classifier on parallel coordinates images to recognize mathematical objects? Probably. Would it be useful? Maybe. For an applied mathematician encountering an unfamiliar dataset, a classifier that said "this looks like a root system of type " could be a real assist.
One way to measure a mathematical idea is to look at how often it reappears in places far from where it started. By that measure, the Inselberg duality is an excellent idea. Started as a visualization trick for correlation analysis. Picked up a general theory of hyperdimensional geometry. Showed up in aircraft separation, robotics, process control, and diagnosis. Turned out to be the natural language for describing regular polytopes in the fourth dimension. Began to make friends with root systems, quasicrystals, codes, lattices, groups, and knots.
Figure 7. The series arc. Six phases, twenty-five articles, one duality.
None of this required anything beyond the original observation. Inselberg saw, in the 1980s, that you could draw the axes side by side and connect the values with a line. He wrote down the convergence-point formula and worked out that projective transformations map cleanly through the duality. And that was enough.
The representation is generative. It keeps finding new uses. Not because parallel coordinates is a universal representation (it is not, as what-didn't-work shows), but because the underlying duality connects two families of geometric objects in a way that turns out to be useful almost everywhere. When a structure has a set of points or lines that obey certain constraints, parallel coordinates lets you see the constraints.
And when you can see the constraints, you can see the structure. That, in the end, is what this series has been about.