Sunday, August 20, 2017

Generalised definition for negative dimensional geometry

In the previous post I discussed a few specialist definitions that allow for negative dimensions, and added my own new one. In this post I propose a slightly different definition of fractal dimension that allows for negative dimensions in a generic way, and fits with the use of negative dimensions in physical quantities. I'll call this the signed-dimension to distinguish it, and because it allows any signed number rather than just positives.

The signed-dimension D of a set should be the length exponent at which that set can be measured.

This is different from the standard notion of dimension, which is the length exponent that part of the set can be measured.

The difference is significant. Normally the dimension of a line is said to be 1, but infinitely long lines can't be measured, only counted. The real-dimension is 0, and the same for infinite planes and volumes. It is only finite extent geometry that has the dimension one usually assigns to it. For example:
  • Line segments have D=1 so their lengths can be measured
  • Disks have D=2 so their areas can be measured
  • Points D=0 so they can be counted. 
What about the set of integers on the real number line? The quantity we can measure is the number of points per distance, in this case one per unit length. So the measure is 1 unit^-1, so D=-1.

By this reasoning the set of integer vectors in 2D space can be measured in points per square length unit, D=-2 and similarly for higher order spaces. Notice that these densities can't have dimension less than the negative of the topological dimension, which mirrors the case for positive dimension shapes. Notice also the consistency that in positive dimensions, if you measure assuming D is higher you get 0 and assuming D is lower you get infinity. For negative dimensions the same is happening but the other way around. For instance the number of integer vectors in 2D space per length unit is infinity.

So negative dimensions refer to densities as positive ones refer to quantities. It is somewhat analogous to contravariant (e.g. vectors) and covariant quantities (covectors or one-forms) in physics. But negative dimensions don't have to just be point-like. For instance a 2D grid of lines spaced a metre apart can be measured as two metres of line per square metre, which is 2m^-1. Nor do they have to take on integer dimensions, for example a set of points equidistant along the Koch curve can be measured as the number of points per length^1.26. Another way to think about negative dimensions is that they represent extrapolations of a replacement method, and positive dimensions represent interpolations:

As you can see from the last two, if the replacement scheme acts on lines then the negative dimension version reflects from the 1D position. The replacement scheme can also act on fractal geometry:
Here the negative dimension version reflects from the 1.26D position of the Koch snowflake. In essence the negative dimension geometry is measuring the density of Koch curves within the infinite extent version of the Koch fractal on the right side.

For replacement schemes, let's call the initial shape in the iterations the tile. Then finite geometry uses the usual Hausdorff dimension:
and the generalised real-dimension adds an extra term:
which is a 0 term on finite geometry since the extrapolate replacement count is 1.

This shows that geometry can be split into a positive dimension (quantity) part and a negative dimension (density) part, and so it is still possible for shapes of infinite extent to have positive dimension, for example an infinite row of disks measures the disk area per metre, so has dimension 1.

In the real world the examples are more stochastic but the definition applies equally. For instance, if you were to measure a country's coastguard capability, you might try to measure the number of coastguards per kilometer of coastline, but this will change depending on what length ruler you use to measure each kilometer. Instead the coastguards can be measured in number per km^1.3 (i.e D=-1.3) where 1.3 is the fractal dimension of a representative piece of the coastline. Just as real coastline is not fractal to infinitely small scales, the coastguards do not cover infinitely large extents. Both are just approximations in the real world.

Aside from the Hausdorff dimension, there are several other dimension measures (box-counting/Minkowski dimension, information dimension, correlation dimension), there are all generalised in the so called q-dimension formula, normally written:
where mu is the natural measure, which is 1/n when that cell of width epsilon is occupied, and n is the number of occupied cells in total.

We can modify this to give a generalised version of the signed-dimension:
where the positive dimension component is:
and the negative dimension component is:
and where:

q=0 is the box counting/Minkowski dimension, q=1 is the information dimension and q=2 is the correlation dimension.

In all cases we can either write the dimension of a set directly, e.g. -0.26D, or leave it in its component form, e.g. 1-1.26D for clarity.

The following post shows many example shapes that are not well represented by just their positive dimension.


  1. I don't like this negative dimension stuff, partly because it breaks monotonicity: if E \subset F then dim E \lessorequal dim F. See Falconer's "Fractal Geometry" 2nd Edition p41 for further desirable properties of dimensions.

  2. Hi Claude thanks for the link. For reference, the properties are: monotonicity, stability, countable stability, geometric invariance, Lipschitz invariance, countable sets, open sets and smooth manifolds.

    The positive dimension D+ satisfies all those properties, and the negative dimension D- satisfies a negative version of them, e.g. if E \in F then dim E >= F. The combined 'signed dimension' does not satisfy all of those properties, but in defense of this:
    1. a dimension satisfying these properties could never be negative, so it is a requirement of a real-valued dimension that you need to have different desirable properties.
    2. I mention in the final sentence that the signed-dimension can be left in its component form, e.g. this shape is 1-1.26D and so the single signed value -0.26D is just an abbreviation. But each component individually does satisfy all the dimension properties you cited.
    3. The single signed dimension does retain what I think is the most important property: the dimension of the Minkowski sum <+> of the sets equals the sum of the dimension of the sets. i.e. Dim(a <+> b) = Dim(a) + Dim(b)