Sources and influences¶

The library compresses a few small books' worth of advice, and each piece has a name.

Five ideas do the work, each with a source:

• The range frame is Tufte's (The Visual Display of Quantitative Information): a frame that shows nothing becomes two spines that show each variable's minimum and maximum.
• Direct labels are Doumont's (Trees, maps and theorems): a legend sends the reader on a round trip between line and key, and a label at the line's end deletes the detour. The label placement is the exact least-squares optimum under no-overlap constraints, re-solved on every draw via the pool-adjacent-violators algorithm.
• The round-number ticks come from Talbot, Lin and Hanrahan's extended Wilkinson algorithm, through mizani's breaks, computed from the data on screen rather than the view limits. How many of them an axis gets is the paper's density term, which mizani leaves out: a target gap between labels in physical units, here tick-label heights, so the count follows the axis's length and the labels' size. The frame modes keep that literature's vocabulary: "nice" numbers (1, 2 or 5 times a power of ten) are Heckbert's (Graphics Gems, 1990), and "loose" is the paper's word for bounds that enclose the data.
• The colours are Paul Tol's colour-blind-safe schemes, arranged ink-first: a single series stays near-black, and colour enters at series two of the same kind. color="tol:orange" works anywhere matplotlib takes a colour.
• The placement rule for label, a fixed side order with the first clear position taken, is cartography's. Imhof (1975) and Yoeli (1972) ranked the positions around a point, Christensen, Marks and Shieber (1995) made that ranking the standard objective, and Vega's label transform (Kittivorawong, Moritz, Wongsuphasawat and Heer, 2021) runs the same greedy model in production. R's directlabels stacks labels along a line with the quadratic programme that placement.stack solves, and its far.from.others.borders is the automatic anchor this library deliberately does not pick. textalloc (MIT) ran the first experiment, and its overlap tests are the pattern the ink harvesting follows.

vanzelfsprekend also joins a long line of Tufte-in-matplotlib work, and its neighbours deserve direct credit:

• dufte, since merged into matplotx, is the closest kin: the same minimal-ink instinct, and its line_labels first framed label placement as a least-squares problem under minimum-distance constraints. vanzelfsprekend solves that same problem, exactly.
• adjustText, following R's ggrepel, tackles the harder general problem of untangling arbitrary 2-D annotations, which takes iterative approximation. Restricting labels to line ends is what lets vanzelfsprekend place them exactly instead.
• etframes is the original, drawing range frames in matplotlib since 2007.