Once points pile up unreadably, binning them into hexagons or a density surface restores a legible pattern

What it is

When points pile up until direct overlay is unreadable, the two standard escapes both turn the point set into a spatially aggregated surface: hexagonal binning (hexbin) tiles the map into equal-area cells colored by their counts, and a density surface (kernel density estimation, KDE, often rendered as a heatmap) smooths discrete points into a continuous density field with a kernel function. Both eliminate overplotting and turn "counting dots" into "reading a surface". Hexagons beat squares on geometric grounds: a regular hexagon is equidistant from all neighbor centers and tiles most compactly, while a square grid treats corner and edge neighbors at different distances and imports a directional bias.

Why it happens

Binning discretizes a resolution-unlimited point set into a finite set of counting cells, each with a well-defined, readable ceiling; saturation moves from "the pixel cannot fit more" to "the cell holds a countable number", and the display-side ceiling disappears. A density surface goes further: each point is spread and weighted by a kernel, presenting density—the analytical object—directly rather than individual positions, at the cost of individuals vanishing entirely. Both share one new free parameter, the bin size or kernel bandwidth. It sets the grain of the visible pattern: the same data can show different spatial structure under different settings, so all the old aggregation problems—compression, boundary sensitivity—come back wearing a new skin, just as cells instead of stacked dots.

Where it stops holding

On small samples, density surfaces oversmooth—a handful of points renders as a large "hotspot". At high zoom, binning works against the reader (one or two points per cell) and the view should fall back to per-point display. Density surfaces truncate at map edges, undercounting true density near boundaries. Neither form shows individual positions, so exact queries ("where is this point?") must fall back on the underlying point data. And because bin size and bandwidth shape the pattern, they must be disclosed to the reader; otherwise a parameter artifact is mistaken for a property of the data.

Applying it

• Default the dense view of an interactive map to hexbin or a density surface, and switch back to individual points automatically as zoom increases and density falls.
• State in the legend what a unit means (count per cell or density units) and disclose the current bin size or bandwidth.
• Verify by running the same data at two or three bin sizes and checking whether the main pattern holds; any region whose pattern changes shape dramatically must carry a "scale-dependent" caveat in conclusions.

Related

• Same group: U12.03.1 Point maps systematically undercount in high-density regions because of overplotting · U12.03.2 Aggregation into clusters hides within-cluster distribution differences · U12.03.3 Cluster markers must encode how many points they contain · U12.03.5 Zoom-dependent aggregation radii make the same data support different conclusions
• Nearby: U12.03.5 Zoom-dependent aggregation radii make the same data support different conclusions · U12.03.1 Point maps systematically undercount in high-density regions because of overplotting
• Search terms: hexbin · kernel density estimation · heatmap