Outer Box Dimension: Theory & Applications

• Outer box dimension is the upper Minkowski dimension defined as the limsup of the logarithm of covering numbers over the negative logarithm of the scale.
• It is central to projection theory, hyperbolic dynamics, and fractal analysis, providing insights into capacities, energy estimates, and invariant graph dimensions.
• The concept extends from bounded Euclidean sets to arbitrary sets via generalized definitions, influencing studies of self-similar attractors, orbital sets, and thermodynamic formalism.

Outer box dimension usually refers to the upper box dimension, or upper Minkowski dimension, of a set, defined from the asymptotic behavior of covering numbers at small scales. In the cited literature this notion is used for bounded subsets of Euclidean space and for sets viewed through explicitly chosen metrics, including the Euclidean metric on $\mathbb R^n$ and the spherical metric on $\partial \mathbb H^n$. It appears in projection theory, hyperbolic-group orbital sets, inhomogeneous attractors, invariant graphs, and level sets of generic Hölder functions. A later extension, the generalized upper box dimension, is defined for arbitrary sets and coincides with the classical upper box dimension on bounded sets (Iommi et al., 2018, Fraser, 2024, Wang et al., 1 Oct 2025).

1. Terminology and basic definition

For a bounded set $E\subseteq \mathbb R^n$, the upper box dimension is defined from covering numbers. Writing $N_r(E)$ or $N_\delta(E)$ for the least number of sets of diameter $r$ or $\delta$ needed to cover $E$, one has

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$

The lower box dimension is obtained by replacing $\limsup$ with $\partial \mathbb H^n$0. When the two agree, their common value is the box dimension (Falconer, 2019, Baker, 2021, Fraser, 2024).

Several papers identify this upper box dimension with the notion often called outer box dimension. One paper states that in its setting upper box dimension is exactly the usual upper Minkowski dimension and is “the same notion often called outer box dimension in some contexts,” while another explicitly says that “outer box dimension” means the upper box dimension of an inhomogeneous attractor (Iommi et al., 2018, Fraser, 2024).

The covering-number definition has equivalent formulations. One paper notes that one may equivalently use the maximal number of disjoint balls of radius $\partial \mathbb H^n$1 with centers in the set, and another defines the upper box dimension by the maximal number of pairwise disjoint closed $\partial \mathbb H^n$2-balls with centers in a bounded metric space (Iommi et al., 2018, Schweinhart, 2018). These equivalent formulations make the invariant adaptable to Euclidean, spherical, and more general metric settings, provided the metric is fixed in advance.

A basic qualification is that the classical definition is naturally tied to bounded sets. The generalized upper box dimension was introduced precisely to remove this restriction: for bounded sets it agrees with the classical upper box dimension, but it is defined for arbitrary sets by means of the Assouad spectrum and upper spectrum (Wang et al., 1 Oct 2025).

2. Capacity profiles and projection theory

In projection theory, the ordinary upper box dimension is often not the correct quantity controlling the typical dimension of orthogonal projections. A capacity-based reformulation introduces, for $\partial \mathbb H^n$3 and $\partial \mathbb H^n$4,

$\partial \mathbb H^n$5

together with the associated energy and capacity

$\partial \mathbb H^n$6

The lower and upper $\partial \mathbb H^n$7-box dimension profiles are then defined by

$\partial \mathbb H^n$8

For $\partial \mathbb H^n$9, these profiles coincide with the ordinary box dimensions, but for $E\subseteq \mathbb R^n$0 they capture the correct almost-sure dimension of projections onto $E\subseteq \mathbb R^n$1-planes (Falconer, 2019).

The projection theorem in this framework states that for every non-empty Borel set $E\subseteq \mathbb R^n$2,

$E\subseteq \mathbb R^n$3

for all $E\subseteq \mathbb R^n$4, with equality for $E\subseteq \mathbb R^n$5-almost all $E\subseteq \mathbb R^n$6. Thus the “typical” projected upper box dimension is governed by the $E\subseteq \mathbb R^n$7-profile rather than by $E\subseteq \mathbb R^n$8 itself (Falconer, 2019).

A complementary line of work shows that Assouad-type regularity can force the ordinary upper box dimension to be preserved under almost all projections. If $E\subseteq \mathbb R^n$9 is bounded and $N_r(E)$0, then for almost all $N_r(E)$1,

$N_r(E)$2

The same conclusion holds under the weaker threshold $N_r(E)$3, and the threshold $N_r(E)$4 is sharp: if the Assouad dimension exceeds $N_r(E)$5, preservation can fail dramatically (Falconer et al., 2019).

Both approaches also quantify exceptional directions. Capacity methods yield Hausdorff-dimension bounds on the set of subspaces where the projected upper box dimension falls below the profile value, while the Assouad-spectrum approach gives exceptional-set bounds strictly below the dimension $N_r(E)$6 of the Grassmannian (Falconer, 2019, Falconer et al., 2019).

3. Orbital sets, Poincaré exponents, and hyperbolic geometry

A major dynamical appearance of outer box dimension occurs for Kleinian orbital sets. If $N_r(E)$7 is a Kleinian group acting on $N_r(E)$8 and $N_r(E)$9 is a non-empty bounded subset of $N_\delta(E)$0, the orbital set is

$N_\delta(E)$1

The principal result is

$N_\delta(E)$2

where $N_\delta(E)$3 is the limit set, $N_\delta(E)$4 is the Poincaré exponent, and the dimension is taken with respect to the Euclidean metric on $N_\delta(E)$5, not the hyperbolic metric (Bartlett et al., 2021).

The formula isolates three distinct sources of complexity: the intrinsic complexity of the seed $N_\delta(E)$6, orbit-growth complexity through $N_\delta(E)$7, and boundary accumulation through $N_\delta(E)$8. The same paper gives examples showing that none of the three terms can be removed in general. For instance, if $N_\delta(E)$9 is generated by a single hyperbolic element and $r$0 is a line segment, then

$r$1

whereas a single parabolic generator with $r$2 a point yields

$r$3

These examples show that the maximum really has three independent entries (Bartlett et al., 2021).

The boundedness hypothesis on $r$4 in the hyperbolic metric is essential. The proof uses the estimate

$r$5

which depends on uniform conformal distortion control for bounded sets. The same paper constructs an explicit counterexample with

$r$6

for which

$r$7

The failure occurs because $r$8 is dense in $r$9, so its closure contains an interval (Bartlett et al., 2021).

A related result concerns parabolic subgroup orbits on the boundary of hyperbolic space. If $\delta$0 is parabolic and $\delta$1 is not fixed by $\delta$2, then

$\delta$3

where the box dimension is computed in the spherical metric on $\delta$4. This ties upper box dimension directly to the critical exponent of the Poincaré series

$\delta$5

through $\delta$6 (Iommi et al., 2018).

4. Pressure, self-similarity, and inhomogeneous attractors

For certain countable-state dynamical systems, outer box dimension can be read from thermodynamic data. If $\delta$7 is an Expanding-Markov-Rényi interval map with branches $\delta$8, the pressure of the geometric potential is

$\delta$9

and there is a critical threshold $E$0 such that $E$1 for $E$2 and $E$3 is finite for $E$4. The boundary of the Markov partition satisfies

$E$5

with equality whenever the box dimension exists (Iommi et al., 2018).

Inhomogeneous iterated function systems supply another central setting. If $E$6 is an IFS of similarities with homogeneous attractor $E$7, condensation set $E$8, and inhomogeneous attractor

$E$9

then the associated orbital set satisfies $$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$0. With $$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$1 denoting the similarity dimension determined by

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$2

the general upper-box estimate is

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$3

Under SOSC, or under the open set condition in Euclidean space, this becomes the exact formula

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$4

(Fraser, 2013, Fraser, 2024).

The expected max-formula can fail when overlaps are present. For the planar self-similar system

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$5

one has

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$6

but if $$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$7 is the reciprocal of a Garsia number, then

$$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$8

For the inhomogeneous Bedford–McMullen-type fractal comb with $$\overline{\dim}_B E=\limsup_{r\to0}\frac{\log N_r(E)}{-\log r}.$$9,

$\limsup$0

These counterexamples show that upper box dimension may depend on the specific IFS and can exceed both $\limsup$1 and $\limsup$2 (Fraser, 2024).

An open problem remains in dimension one: for an inhomogeneous self-similar set $\limsup$3, it is asked whether

$\limsup$4

always holds (Fraser, 2024).

5. Graphs, level sets, and oscillatory geometry

For invariant graphs of hyperbolic skew products, box dimension is frequently given by pressure equations. If

$\limsup$5

is the unique invariant graph of a three-dimensional skew product and $\limsup$6 is the unique solution of

$\limsup$7

with stable-slice dimension $\limsup$8 determined by

$\limsup$9

then in the non-Lipschitz regime the graph satisfies

$\partial \mathbb H^n$00

In the Lipschitz regime the dimension is smaller: for an Anosov base $\partial \mathbb H^n$01, and for a one-dimensional hyperbolic attractor $\partial \mathbb H^n$02 (Díaz et al., 2017).

For generalized affine recurrent fractal interpolation functions, the controlling quantities are the spectral radii of restricted vertical scaling matrices. If $\partial \mathbb H^n$03 satisfies the standing assumptions (A1)–(A4), then the main upper bound is

$\partial \mathbb H^n$04

where $\partial \mathbb H^n$05 is the limit of the spectral radii of the restricted matrices in the $\partial \mathbb H^n$06-th strongly connected component. Under positivity of the scaling functions on the relevant invariant sets, irreducibility and Perron–Frobenius theory yield the exact formula

$\partial \mathbb H^n$07

with

$\partial \mathbb H^n$08

whenever infinite variation occurs on an appropriate basic interval (Jiang et al., 3 Oct 2025).

Upper box dimension of level sets behaves differently from Hausdorff dimension and lower box dimension. For generic $\partial \mathbb H^n$09-Hölder-$\partial \mathbb H^n$10 functions on a compact fractal $\partial \mathbb H^n$11, the upper-box theory is described as measuring how much level sets can spread across the fractal, or how widely the generic function can oscillate on it. If $\partial \mathbb H^n$12 has nice connection type, then there is a dense $\partial \mathbb H^n$13 such that every $\partial \mathbb H^n$14 has the same typical upper-box dimension of level sets,

$\partial \mathbb H^n$15

For the cube $\partial \mathbb H^n$16,

$\partial \mathbb H^n$17

and for the Sierpiński triangle $\partial \mathbb H^n$18,

$\partial \mathbb H^n$19

(Buczolich et al., 2023).

Arithmetic orbit problems furnish another oscillatory setting. For $\partial \mathbb H^n$20 orbits on the circle, if $\partial \mathbb H^n$21, $\partial \mathbb H^n$22, and $\partial \mathbb H^n$23, then

$\partial \mathbb H^n$24

A highlighted corollary states that if $\partial \mathbb H^n$25 is not a Liouville number and $\partial \mathbb H^n$26 is a Liouville number, then any $\partial \mathbb H^n$27 orbit, and hence any $\partial \mathbb H^n$28 set, has

$\partial \mathbb H^n$29

(Baker, 2021).

6. Generalizations, related invariants, and distinctions

The most explicit extension of outer box dimension beyond bounded sets is the generalized upper box dimension

$\partial \mathbb H^n$30

For bounded $\partial \mathbb H^n$31, one has

$\partial \mathbb H^n$32

This generalized dimension retains standard properties such as monotonicity, finite-union stability, bi-Lipschitz invariance, closure invariance, and product subadditivity, and it satisfies

$\partial \mathbb H^n$33

Its modified version recovers packing dimension: $\partial \mathbb H^n$34 Moreover,

$\partial \mathbb H^n$35

(Wang et al., 1 Oct 2025).

A topological analogue arises from persistent homology. For a bounded subset $\partial \mathbb H^n$36 of a metric space, the $\partial \mathbb H^n$37-dimension is defined by the threshold exponent controlling

$\partial \mathbb H^n$38

uniformly over all finite $\partial \mathbb H^n$39. In degree $\partial \mathbb H^n$40, this recovers the MST dimension and equals the upper box dimension. In the plane, if $\partial \mathbb H^n$41 and

$\partial \mathbb H^n$42

then

$\partial \mathbb H^n$43

This gives a nontrivial comparison between upper box dimension and a persistent-homological fractal invariant (Schweinhart, 2018).

A frequent source of confusion is the phrase box space in coarse geometry. The asymptotic dimension of a box space

$\partial \mathbb H^n$44

is a different invariant: it concerns the coarse disjoint union of finite quotients of a group and is measured by large-scale coverings rather than small-scale covering growth. The relevant standard term there is the asymptotic dimension of the box space or box family, not outer box dimension in the upper-Minkowski sense (Finn-Sell et al., 2015).

Taken together, these developments show that outer box dimension is both a classical covering invariant and a nodal point connecting thermodynamic formalism, hyperbolic orbit growth, projection theory, self-similar and inhomogeneous constructions, arithmetic dynamics, generic oscillation phenomena, and more recent spectrum-based generalizations. The recurring pattern is that the invariant is simple at the definitional level, but its exact value is often governed by additional structures: pressure thresholds, Poincaré exponents, capacity profiles, Assouad-type regularity, or spectral radii of transfer-like matrices.