Minkowski dimension

Minkowski dimension is defined by the asymptotic growth of covering numbers as the scale tends to zero. For a bounded set E in Euclidean space, let N(ε) be the smallest number of sets of diameter at most ε needed to cover E. The upper and lower box dimensions are respectively the limsup and liminf of log N(ε) divided by −log(ε); when they agree, their common value is the Minkowski dimension.

Equivalent constructions use a grid of cubes of side length ε or the volume of an ε-neighborhood, often called a Minkowski sausage. Changing between standard reasonable shapes of covering sets does not alter the dimension, although finite-scale numerical estimates can differ substantially.¹

Minkowski dimension reproduces familiar dimensions for sufficiently regular objects: a finite collection of points has dimension 0, a smooth curve has dimension 1, and a planar region with nonempty interior has dimension 2. A self-similar fractal can instead have a fractional value. For the middle-thirds Cantor set, the dimension is log 2 divided by log 3, approximately 0.6309; the Koch curve has dimension log 4 divided by log 3, approximately 1.2619.

The value describes covering complexity rather than visual roughness alone. A set may have the same Minkowski dimension as another set while possessing different geometry, topology, or measure. In particular, positive or full Lebesgue measure does not by itself determine the fine-scale arrangement captured by box counts.²

Minkowski dimension is closely related to Hausdorff dimension but is generally easier to estimate and can be strictly larger. For every bounded set, the Hausdorff dimension is no greater than the lower box dimension, which is no greater than the upper box dimension. Thus a set can have unequal upper and lower box dimensions, in which case it has no single Minkowski dimension even though its Hausdorff dimension is well defined.

Unlike Hausdorff dimension, box dimension is not countably stable: a countable union of sets can have a larger box dimension than any one member. This distinction matters in dynamics, where the box dimension of an attractor or orbit closure may be convenient computationally but can behave differently from the more measure-theoretic Hausdorff dimension.¹

Numerical box-counting is an asymptotic method, so finite data can produce a convincing but misleading straight line on a log–log plot. Grid placement, anisotropic scaling, limited resolution, noise, and the selected fitting interval all affect the estimated slope. The most reliable studies test several scales and report how sensitive the result is to these choices.

The dimension also has less familiar variants. Packing-based formulations, modified box dimensions, and dimensions of graphs are used when ordinary covering counts lose useful information. In fractal geometry, the leading coefficient of neighborhood volume can carry additional information: when a suitable normalized limit exists, it is called the Minkowski content, and its existence is the subject of Minkowski measurability.

Box-counting dimension

A common synonym for Minkowski dimension, based on the scaling of the number of small boxes covering a set.

Upper box dimension

The limsup scaling exponent of covering numbers as the box size approaches zero.

Lower box dimension

The liminf scaling exponent of covering numbers as the box size approaches zero.

Minkowski sausage

The neighborhood obtained by taking all points within a specified distance of a set; its volume can encode Minkowski dimension.

Minkowski content

A normalized limiting neighborhood volume that provides size information beyond the dimension when the limit exists.