A point in Cartesian space becomes a polyline in parallel coordinates. A line becomes a point. This inversion is the foundation of everything that follows.
In Cartesian coordinates, a point sits at the intersection of its values, which limits the display to two or three dimensions. Parallel coordinates draw the axes side by side and connect values with line segments. A 2D point becomes a line; an N-dimensional point becomes a polyline with N−1 segments.
This is a precise mathematical duality, discovered by Alfred Inselberg in the 1980s. Points map to lines, lines map to points, and the mapping preserves geometric structure.
In parallel coordinates, draw two vertical axes separated by a distance. A point in two-dimensional Cartesian space maps to a line segment connecting its first coordinate value on the left axis to its second coordinate value on the right axis.
Drag the point in Cartesian, or drag the dots on the parallel coordinate axes. Both sides update.
Figure 1. A point in Cartesian space (left) maps to a line segment in parallel coordinates (right). The duality is bidirectional: drag either representation and the other follows.
The duality is invertible. Each side is a complete encoding of the other.
Figure 2. Three ways to explore the duality. Click in Cartesian (left) to place a point, and its line appears in PC. Click on a PC axis, then the other, to build a line whose point appears in Cartesian. Click between the PC axes to place a convergence point, and its line appears in Cartesian. All placed items are draggable. Toggle "Show fan lines" to see sample segments passing through each convergence point.
The mapping generalizes to any number of dimensions. A point in N-dimensional space has N coordinate values. In parallel coordinates, draw N vertical axes. The point maps to a polyline that visits each axis at the height of the corresponding coordinate value.
Figure 3. Multiple points in N dimensions. Each point becomes a polyline visiting N axes. Toggle the number of dimensions to see the representation scale.
The duality runs deeper than point-to-polyline. Consider a set of points that all lie on a single line in Cartesian space. Each point maps to a line segment in parallel coordinates. These segments all pass through a single point, called the convergence point or indexed point.
For a Cartesian line defined by the equation , where is the slope and is the intercept, the convergence point in parallel coordinates (with unit axis spacing) is located at:
The horizontal position depends only on slope; the vertical position encodes the intercept. Rotating a Cartesian line (slope change) translates the convergence point horizontally; translating it (intercept change) moves the convergence point vertically.
Drag the control points in Cartesian, or drag the blue dots on the PC axes. The convergence point (red) follows.
Figure 4. Points on a Cartesian line (left) produce line segments in parallel coordinates (right) that all pass through a single convergence point (red dot). Drag from either side to explore how slope and intercept map to the convergence point's position.
The horizontal position of the convergence point reveals the nature of the relationship between the two variables. This is a direct consequence of the formula: the denominator is , so the sign and magnitude of the slope determine where the point falls.
Figure 5. How the convergence point moves as slope changes. Negative slopes place it between the axes. Slope of 1 sends it to infinity (parallel segments). Slopes greater than 1 place it left of the left axis.
Negative correlation (m < 0): convergence falls between the axes. The line segments cross, forming an X-pattern. At m = −1, the convergence point sits exactly at the midpoint.
No correlation (m = 0): convergence is on the right axis. All segments fan from different heights on the left axis to the same height on the right.
Positive correlation (0 < m < 1): convergence is to the right of the right axis. Segments run roughly parallel.
Perfect positive (m = 1): segments are exactly parallel. The convergence point is at infinity. This is the projective case.
Steep positive (m > 1): convergence is to the left of the left axis. Segments cross again, but the crossing pattern differs from the negative-slope case.
At m = 1, the formula divides by zero. Geometrically, all the line segments in parallel coordinates become exactly parallel and the convergence point sits at infinity. The duality lives in projective space, where points at infinity are first-class citizens.
As slope approaches 1, the convergence point races off to the left or right. At m = 1 it wraps around to the point at infinity. Projective geometry handles this seamlessly.
When a dataset contains many points, the parallel coordinates view fills with polylines. The structure of these polylines encodes the structure of the data. Points that are close in Cartesian space produce polylines that are close in parallel coordinates. Linear relationships produce convergence patterns. Clusters produce bundles.
Figure 6. Different data structures produce different polyline patterns. Random data fills the space. Correlated data produces crossing patterns. Clusters produce bundles. Points on a line converge.
In N dimensions, a line is determined by two points. In parallel coordinates with N axes, those two points produce two polylines. The polylines intersect at exactly N−1 points, one between each pair of adjacent axes. These N−1 intersection points are the indexed points that fully characterize the line.
A line in N-dimensional space has 2(N−1) degrees of freedom. The N−1 indexed points each have 2 coordinates, giving 2(N−1) parameters. The encoding is exact. No information is lost.
The two highlighted polylines define a line in 5D. Their intersections (red dots) are the 4 indexed points characterizing that line.
Figure 7. Two points in 5D define a line. Their polylines intersect at 4 indexed points (red dots) between adjacent axes, fully encoding the line's direction and position.
The duality between points and polylines, lines and convergence points, is the foundation for everything else in the series — cluster detection through polytope visualization. It's an invertible mathematical transformation, not a metaphor.