Geometric measure theory studies fractal measures and dimensions. We outline here a standard method for building outer meaures from premeasures, which guides the standard construction of Hausdorff measures and dimension. These notes are biased to put a spotlight on Hausdorff dimension as a principal quantity of fractal geometry. This is despite the established importance of other fractal measures/dimensions and their established connections with AIT.
We begin by reviewing premeasures and their associated outer measures. For more details, refer to Reimann, 2008 and Rogers, 1970.
In the setting of Cantor space, it is customary to consider a premeasure on 2ω to be any non-negative mapping defined on the cylinder setsρ:B→[0,+∞]. Of course, such maps are in one-to-one correspondence with maps defined on the finite binary strings ρ:2<ω→[0,+∞]. We may freely interchange between these two types by identifying any σ∈2<ω with its cylinder set [σ]∈B.
In metric spaces, there is a standard, general method (introduced as “Method II” in Rogers (1970)) to induce an outer measure from a given premeasure.
We call Hρ the Method II outer measure associated to the premeasure ρ. And for each δ>0, we also call Hδρ the ρ-dimensional δ-content measure.
An outer measure μ on a metric space is metric if μ(E∪F)=μ(E)+μ(F) whenever E and F are a positive distance apart. Every Method II outer measure is metric: if E and F are a distance d>0 apart, then no set of diameter less than d meets both, so Hδρ(E∪F)=Hδρ(E)+Hδρ(F) for every δ<d. By Carathéodory’s criterion, an outer measure is metric if and only if every Borel set B is μ-measurable, meaning μ(E)=μ(E∩B)+μ(E∖B) for every E. Every Method II outer measure is metric Rogers, 1970.
In Rogers (1970), C. Rogers also defines an alternative “Method I” which differs from Method II in that it does not enforce any upper bound on the diameter of the cover elements used in the infimum:
HIρ(X):=(Ci)i∈ω⊆Cinf{i∑ρ(Ci):i⋃Ci⊇X}.
This outer measure is also sometimes denoted H∞ρ. While this method also produces an outer measure from a premeasure ρ which agrees on null sets with Hρ, Rogers points out that only the results of Method II can be guaranteed to be additive on sets which are separated by positive distance Rogers, 1970. We proceed with Method II as our preferred way to produce outer measures from premeasures.
Cantor space: 2ω, is a metric space when imbued with the standard metric,
d(x,y):={2−N0if x=y and N=min{n∈ω:x(n)=y(n)},if x=y.
Note that this metric is compatible with the product topology generated by the cylinder sets in B.
An important class of outer measures on any metric space (Ω,d) are the Hausdorff outer measures, which are those Method II outer measures induced from premeasures whose values depend only on the diameter of the set. First, declare a dimension function (a.k.a., gauge) to be any non-negative, non-decreasing, continuous-on-the-right function h defined on all non-negative real numbers and satisfying h(t)=0⟺t=0. One may associate to any dimension function h its Hausdorff premeasure defined on all X⊆Ω as follows:
ρh(X):=(h∘diamd)(X).
Denote by Hh:=Hρh the Hausdorff outer measure associated to h.
In Cantor space, ρh(σ)=h(2−len(σ)) for any σ∈2<ω. So, any Hausdorff premeasure on 2ω is length-invariant, in that ρh(σ) only depends on the length of σ.
Later, we will briefly focus on convex Hausdorff premeasures ρ over Cantor space, which satisfy:
ρ(σ)≤ρ(σ⌢0)+ρ(σ⌢1),
for all σ∈2<ω. It is easy to see that all length-invariant, convex Hausdorff premeasures over 2ω are also strongly convex, meaning for all σ∈2<ω and i∈{0,1},
ρ(σ⌢i)≥2ρ(σ).
Essential to fractal geometry is the family of s-dimensional Hausdorff premeasuresρs:=ρhs, where hs:t↦ts is the dimension function associated to s≥0. If δ>0, let Hδs:=Hδρs denote the s-dimensional Hausdorff δ-content, and Hs:=Hρs the s-dimensional Hausdorff outer measure.
For any outer measure μ on a metric space (Ω,d), we say a set X⊆Ω is μ-null if μ(X)=0. Any Hs-null X is also called s-null. And, if μ and ν are outer measures on Ω, then ν is said to be absolutely continuous with respect to μ, written ν≪μ, if every μ-null set is also ν-null. A set X⊆Ω is μ-conull if Ω∖X is μ-null. More generally, a subset Y⊆X is μ-conull in X if X∖Y is μ-null.
Suppose Hs(X)<∞ for some s≥0. This implies that X is s+-null for any s+>s too. Alternatively, if Hs(X)>0 for some s≥0, then it must be that Hs−(X)=∞ for all 0≤s−<s. So, the shape of the function s↦Hs(X) is either constant with value 0 or ∞, or a step from ∞ to 0 (where, at the critical value of s where the step occurs, Hs(X) could take on any value in [0,∞]). This motivates the following definition for the Hausdorff dimension of a set XHausdorff, 1919Falconer, 1990:
Consider now how some sets fall between the cracks of the power functions hs: a set of Hausdorff dimension s could have Hs(X)∈{0,+∞} (i.e., be s-null or have infinite Hs-measure). In either case, no power function hs′ assigns X a positive, finite measure. Still, it is possible that some dimension function does: call a dimension function h an exact gauge for X if 0<Hh(X)<+∞. In particular, if the gauge hs is exact for a set X, then it is common to call X and s-set.
And exact gauges still relate back to Hausdorff dimension through their relationship with power laws.
If X is a set for which there exists an exact gauge h for X with liminft→0+logtlogh(t)=dimH(X), let us consider X to be Hausdorff gauge-regular. In particular, every s-set is Hausdorff gauge-regular, as witnessed by hs.
Over Euclidean space, a second standard fractal dimension is built from the covering numbers:
While Tricot (1982) originally introduced packing dimension through packing measures, the formula above is its equivalent characterization as the modified upper box-counting dimension Falconer, 1990. We always have that dimH(X)≤dimP(X), and dimP(X)≤dimB(X) whenever X is bounded. Following Fiedler & Stull (2024), a set X⊆Rm is said to be weakly regular if dimH(X)=dimP(X).
An outer measure inherits Hausdorff dimension from the sets that carry its mass.
Intuitively, the lower Hausdorff dimension is small when at least some of its mass sits on a set of small dimension; whereas the upper Hausdorff dimension is small when all of its mass does. These definitions are usually stated for finite Borel measures over RmFalconer, 1997.
Notice that every μ-conull set E has μ(E)=μ(Ω)>0, so we always have dimH(μ)≤dimH(μ).
For an outer measure μ on Ω and a set X⊆Ω, let the restriction of μ to X be the outer measure μ└X:E↦μ(E∩X). We can also relate the dimension of μ└X back to that of X. Fix X⊆Ω with μ(X)>0. Notice that a set E is (μ└X)-conull exactly when E∩X is μ-conull in X, so
dimH(μ└X)=inf{dimH(Y):Y⊆X is μ-conull in X}=inf{dimH(X∖N):N⊆X is μ-null}≤dimH(X).
Equality here holds exactly when deleting a μ-null set N from X never lowers its Hausdorff dimension. Let’s call a set XHausdorff measure-regular if there exists a metric outer measure μ on Ω with 0<μ(X)<∞ and dimH(μ└X)=dimH(X).
The following proposition gives two sufficient conditions for this:
Stull (2022) assumes these two conditions, for a metric outer measure μ with 0<μ(X)<∞, in his sufficient condition for the existence of optimal Hausdorff oracles.
Note that while the first condition actually proves equality for lower Hausdorff dimension as well, the second condition does not control the lower dimension. To be concrete, suppose s>0, that Hs(X)>0, and x∈X. Then μ:=Hs+δx (where δx is the point mass at x) satisfies the second condition while dimH(μ└X)=0.
Falconer, K. (1990). Fractal geometry - mathematical foundations and applications. (p. I–XXII, 1–288). Wiley.
Tricot, C. (1982). Two definitions of fractional dimension. Mathematical Proceedings of the Cambridge Philosophical Society, 91(1), 57–74. 10.1017/S0305004100059119
Fiedler, J. B., & Stull, D. M. (2024). Universal Sets for Projections.
Falconer, K. (1997). Techniques in Fractal Geometry. Wiley.
Stull, Donald. M. (2022). Optimal Oracles for Point-To-Set Principles. In P. Berenbrink & B. Monmege (Eds.), 39th International Symposium on Theoretical Aspects of Computer Science (STACS 2022) (Vol. 219, p. 57:1-57:17). Schloss Dagstuhl – Leibniz-Zentrum für Informatik. 10.4230/LIPIcs.STACS.2022.57