Band width is a length encoding; one scale for the whole diagram

What it is

A Sankey encodes flow with the transverse width of a band, a length channel. “One scale for the whole diagram” means that one linear conversion applies within the comparison scope: equal flows receive equal widths and widths can add to node totals. If a stage, thin flow, or local region is magnified independently, readers can no longer compare widths across the diagram or audit conservation. Colour, opacity, and path length may support categorization and tracing but should not silently alter the quantity represented by width.

Why it happens

A shared coefficient makes width = k × flow hold for every quantitative band, so width ratios equal flow ratios and incoming widths can be compared with outgoing and balance terms. Nonlinear widths, piecewise coefficients, and local fitting break both proportionality and additivity. Curvature does not alter the theoretical transverse width, but joins, antialiasing, and scaling can change visible thickness. Validate source geometry or a cross-section perpendicular to the centreline rather than measuring an arbitrary diagonal edge.

Studying it

Engineering tests can feed known flow ratios through the pipeline, inspect pre-layout widths and final vector paths, render at multiple sizes, and check split and merge conservation. Reader studies can compare a shared linear scale, local rescaling, explicitly labelled panel scales, and indicative minimum widths on ratio, ranking, node-audit, and cross-stage tracing tasks. Tolerance should follow rendering conditions and decision risk rather than one fixed percentage. Browser zoom, responsive layouts, screenshots, print, and high-contrast modes belong in the test matrix.

Where it stops holding

“Whole diagram” does not force every independent Sankey to share a scale. Panels need a common width scale when readers must compare absolute flows across them. If panels support only within-panel composition, each may use a clearly labelled scale while explicitly ruling out cross-panel width comparison. Opening a detail view and rescaling every band within that view creates a new comparison scope rather than a local cheat; signal the scale change and preserve return. An indicative minimum width may aid detection or selection, but it must be distinguished from quantitative width and node accounts must retain true values.

Applying it

• Route every quantitative band through one linear width function and prohibit local fitting, logarithmic widths, or unlabelled minimum widths in the same diagram.
• Supply a width reference, unit, and node values. Mark multiple panels as shared-scale or independent-scale and let readers lock a scale for comparison.
• Separate visible path from a larger hit target, adding keyboard selection and tabular access for thin flows instead of exaggerating their quantitative width.
• Regression-test known ratios and balanced nodes, then test magnitude and tracing tasks at target output sizes. Investigate data, layout, and rendering separately when ratios drift.

Related

• Same group: U2.12.1 Sankey diagrams encode conserved flows as band width · U2.12.2 Node ordering and crossing count decide readability · U2.12.3 With huge magnitude gaps, small flows vanish · U2.12.4 Backflow and loops disqualify the Sankey
• Nearby: U1.11.3 Only charts sharing a scale can be compared on magnitude · U2.15.2 All facets must share the same scale to be comparable
• Search terms: Sankey width scale · linear width encoding · shared scale · flow additivity