Box-Counting Dimension, Explained

After reading this you will be able to cover any shape with a grid, count the boxes it touches at several scales, and read off a non-integer dimension from the slope of a straight line on log-log paper.

What box-counting measures, and one example

A line has dimension 1. A filled square has dimension 2. A point has dimension 0. These are the dimensions you learned as a child, and they are all whole numbers. Box-counting dimension extends that idea to shapes that fall between the integers.

Here is the hook. Take a Koch curve: start with a straight segment, replace its middle third with two sides of a small triangle, then repeat on every segment forever. The result is a wiggly curve of infinite length that still fits inside a finite box. It is too crinkly to be a plain line, but it does not fill any area. Its box-counting dimension turns out to be \log 4 / \log 3 \approx 1.262. That number sits strictly between 1 and 2, and it captures exactly how the curve fills space more thoroughly than a smooth line but less than a region.

The method behind that number is plain counting. Lay a grid of square boxes of side \varepsilon over the shape. Count how many boxes the shape touches. Call that N(\varepsilon). Shrink the boxes and count again. The way N(\varepsilon) grows as \varepsilon shrinks tells you the dimension.

The intuition before the formula

Think about the shapes you already understand. Cover a straight segment of length 1 with boxes of side \varepsilon = 1/4. You need about 4 boxes. Halve the box size to 1/8 and you need about 8. Each time you halve \varepsilon, the count doubles. Count grows like 1/\varepsilon to the first power.

Now cover a filled unit square with the same boxes. At \varepsilon = 1/4 you need 16 boxes. At 1/8 you need 64. Halving \varepsilon multiplies the count by 4, so the count grows like 1/\varepsilon to the second power.

The exponent is the dimension. A line gives exponent 1, a region gives exponent 2. A fractal gives something in between because it fills space in a way that is neither.

The formula and how to fit it

Write the counting relationship as a power law:

Here N(\varepsilon) is the number of occupied boxes at box size \varepsilon, C is a constant that depends on the shape's overall size, and D is the box-counting dimension you want. The minus sign appears because smaller boxes mean a larger count.

Take the logarithm of both sides to turn the power law into a straight line:

Now plot \log(1/\varepsilon) on the horizontal axis and \log N(\varepsilon) on the vertical axis. The points fall near a line whose slope is D. The intercept is \log C and you can ignore it. Any base of logarithm works as long as you use the same base on both axes, because the base cancels in the slope.

To get D from several points, fit a line by least squares. For points (x_i, y_i) with x_i = \log(1/\varepsilon_i) and y_i = \log N(\varepsilon_i), the slope is:

where \bar{x} and \bar{y} are the means of the two coordinates. That single number is the estimate you report.

A worked example on the Koch curve

The demo runs on a Koch curve with the field defaults. Work the numbers by hand using the natural scaling of the construction. At each stage of the Koch curve there are 4 copies of the previous shape, each scaled down by a factor of 3. That is exactly what a clean box count sees.

Reading and interpreting the slope

The slope you fit is the whole answer. Compare it against the two anchors: 1 for a smooth curve, 2 for a filled region. The Sierpinski triangle sits at \log 3 / \log 2 \approx 1.585 because each halving of the box triples the occupied count. A measured coastline usually lands between about 1.02 and 1.30 depending on how rugged it is. Britain's west coast is often quoted near 1.25.

The table below shows the ideal counting behaviour for four shapes. Read each row as the factor by which the box count multiplies when you halve \varepsilon, and the dimension that factor implies.

The Koch factor 2.520 comes from 2^{1.262}. Any factor between 2 and 4 gives a dimension between 1 and 2, which is the fractal band.

Common mistakes

The estimate is only as honest as the range of scales you fit. Watch for four traps.

Fitting the flat ends

At very large boxes the whole shape sits in one box, so N = 1 and the slope goes to 0. At very small boxes, once every box holds at most one pixel, doubling the resolution just doubles the count and the slope drifts toward 2. Fit only the straight middle of the log-log plot.

Too few scales

Two points always define a line, so two scales give a slope with no way to check it. Use at least four or five scales and look at how tightly they line up.

Grid placement

Shifting the grid by half a box can change N(\varepsilon) by 10 to 20 percent at coarse scales. Serious estimates average over several grid offsets.

Confusing dimension with roughness you can see

A dimension of 1.26 does not mean the curve is 26 percent longer. It describes a scaling rate, not a length. Fractal length grows without bound as \varepsilon shrinks.

Related tools

The same self-similar fractals appear all over this site, built by very different rules. The Chaos Game grows the Sierpinski triangle from random jumps, and Pascal's Triangle mod n reveals the same shape by colouring remainders. Zoom into the Mandelbrot Explorer or slide the parameter in the Julia Set Explorer to see boundaries whose dimension is exactly 2. For a curve that fills a square completely, so its box dimension is 2, watch the Hilbert Curve Explorer. Strange attractors with fractal cross-sections show up in the Lorenz Attractor and the Hénon Map Attractor, and the Magnetic Pendulum Fractal shades a fractal basin boundary you could box-count yourself.

Frequently asked questions

Related tools

3D Geometry Board

Points, lines, planes, spheres and polyhedra on an orbiting camera — slice solids with planes, unfold nets, drop the Dandelin spheres into a cone and drag everything.

Bézier Curve Playground

Drag control points and watch De Casteljau’s algorithm construct a Bézier curve step by step.

Central Limit Theorem Demo

Sample from any lopsided distribution and watch the averages pile up into a bell curve.

Conway’s Game of Life

Paint cells and watch Conway’s cellular automaton breed gliders, guns and oscillators.

Dynamic Geometry Board

Construct with points, lines, circles and transformations, then drag anything — the construction stays true, and a checker tells you which relations are theorems.

Fantasy Map Generator

Roll a seed and get a continent: coastlines, mountains, rivers, biomes and named towns, ready for your campaign.