Four Checks Before Trusting a Dimension Calculation
Hausdorff-dimension arguments often involve simple arithmetic—raising a cover size to a power, taking an infimum, or solving a scaling equation. The danger is that these operations encode different choices: a ball’s radius is not its diameter, the best cover may depend on the scale, a similarity equation does not account for overlaps, and a dimension value does not specify the measure at that value. We isolate these distinctions and test each with a worked calculation.
Let \((X,d)\) be a metric space and \(E\subseteq X\). For \(s>0\) and \(\delta>0\), define the \(s\)-dimensional Hausdorff content at scale \(\delta\) by
The infimum is over countable covers by subsets of \(X\); empty members can be discarded. We use \(0^s=0\) for \(s>0\). The Hausdorff measure is \(\mathcal{H}^s(E)=\lim_{\delta\downarrow0}\mathcal{H}^s_\delta(E)\), a limit that exists in \([0,\infty]\) because the contents do not decrease as \(\delta\) decreases. In this tutorial the dimension convention is
This convention uses only exponents for which the measure has been defined. For the empty set the set inside the infimum is all \(s>0\), so \(\dim_H\varnothing=0\). We will use the established facts that Hausdorff measure is countably subadditive and Hausdorff dimension is countably stable: \(\dim_H(\bigcup_n E_n)=\sup_n\dim_H E_n\).
Worked Example: A Ball’s Radius Is Not Its Cover Diameter
Worked Example: Recovering the Length of an Interval
Use \(N\) equal closed intervals to cover \(E=[0,1]\subset\mathbb{R}\), where \(\mathbb{R}\) has its usual distance. The \(k\)-th interval is \([(k-1)/N,k/N]\), for \(k=1,\ldots,N\). Each interval has diameter \(1/N\), so the total cost in the definition of \(\mathcal{H}^1_\delta\) is
When \(1/N\le\delta\), this is an admissible cover, giving \(\mathcal{H}^1_\delta([0,1])\le1\). The established interval formula for Hausdorff measure gives the reverse bound, so \(\mathcal{H}^1([0,1])=1\). The same intervals are balls in \(\mathbb{R}\) of radius \(1/(2N)\), but their diameter remains \(2(1/(2N))=1/N\). A mistaken radius-based cost would be \(N(1/(2N))=1/2\), which contradicts the interval formula. Check: substituting \(N=4\), each interval has radius \(1/8\) and diameter \(1/4\); the correct total cost is \(4(1/4)=1\), whereas the radius mistake gives \(4(1/8)=1/2\).
More generally, a ball of radius \(r\) in a metric space has diameter at most \(2r\), not necessarily exactly \(r\). Using \(r^s\) instead of \((\operatorname{diam}U)^s\) changes a cover’s cost. In Euclidean settings the bound \(\operatorname{diam}B(x,r)\le2r\) can give a valid estimate with a factor \(2^s\), but it does not make radius and diameter interchangeable in an exact calculation.
Scale Limits and Infima Must Stay in Their Order
At each fixed \(\delta\), the infimum in \(\mathcal{H}^s_\delta(E)\) selects the least cost among covers whose members have diameter at most \(\delta\). When \(\delta\) changes, the admissible covers change too. The definition takes the infimum at each scale first, then the limit as scale tends to zero. An estimate for one chosen cover, or for a sequence of covers whose scale changes, does not by itself justify reversing those operations.
Worked Example: A Limit and an Infimum Can Disagree
For positive integers \(n,k\), set \(a_{n,k}=1\) if \(k\le n\), and \(a_{n,k}=0\) if \(k>n\). For each fixed \(n\), there is a \(k>n\), so \(\inf_k a_{n,k}=0\). Therefore \(\lim_{n\to\infty}\inf_k a_{n,k}=0\). But for each fixed \(k\), every \(n\ge k\) gives \(a_{n,k}=1\), so \(\lim_{n\to\infty}a_{n,k}=1\); hence \(\inf_k\lim_{n\to\infty}a_{n,k}=1\). Thus
This computation is not itself a Hausdorff-measure calculation; it shows why the proposed exchange is not a valid general rule. In a Hausdorff calculation, the index \(k\) can represent a cover choice, and the best choice can vary with the scale \(1/n\). To conclude a uniform bound, one needs an estimate that applies to every sufficiently small scale, not just a favorable cover at each selected scale. Check: at \(n=3\), choosing \(k=4\) gives \(a_{3,4}=0\), while for any fixed \(k\), taking \(n\ge k\) gives \(a_{n,k}=1\); the two operations select different information.
A sound covering argument therefore keeps the order explicit: fix the exponent, take the infimum over all covers at scale \(\delta\), and then let \(\delta\downarrow0\). If an argument instead takes a limit along a particular family of covers, it has established an upper bound only when those covers are admissible and their costs have the claimed limit. A lower bound must address every admissible cover, often using a separate measure or covering estimate.
Similarity Dimension Does Not Ignore Overlaps
For similarities with contraction ratios \(r_1,\ldots,r_q\), the similarity dimension is the nonnegative solution \(s\) of \(\sum_{i=1}^q r_i^s=1\), when such a solution exists. This equation records the contraction ratios, not how the images sit relative to one another. The similarity dimension agrees with the Hausdorff dimension under suitable separation hypotheses, such as the open set condition; without such a hypothesis it need not.
Worked Example: Two Distinct Maps with the Same Fixed Point
On \(\mathbb{R}\), consider the contractions \(S_1(x)=x/2\) and \(S_2(x)=x/3\), both mapping \([0,1]\) into \([0,1]\). Their similarity equation is \((1/2)^s+(1/3)^s=1\). The left side is continuous and strictly decreasing in \(s\), equals \(2\) at \(s=0\), and equals \(1/2+1/3=5/6<1\) at \(s=1\). Thus there is a unique solution \(s\) with \(0<s<1\).
The invariant compact set is \(K=\{0\}\): both maps fix \(0\). To check uniqueness, any point in the invariant set obtained after \(n\) successive maps has the form \(S_{i_1}\circ\cdots\circ S_{i_n}(x)\) for some \(x\in[0,1]\). Its value lies in \([0,2^{-n}]\), since every contraction ratio is at most \(1/2\). As \(n\to\infty\), these intervals shrink to \(\{0\}\), so the attractor is the singleton. A singleton has Hausdorff dimension \(0\): for every \(t>0\), it can be covered by one interval of diameter \(\delta\), whose \(t\)-cost is \(\delta^t\to0\). The similarity equation gives \(s>0\), whereas \(\dim_H K=0\). The nested images overlap, and the equation alone therefore overstates the dimension. Check: at exponent \(1\), the ratio sum is \(1/2+1/3=5/6\); at exponent \(0\), it is \(1+1=2\), confirming that its root is strictly between \(0\) and \(1\).
The diagnostic is not “never use the similarity equation.” It is to verify that the separation hypotheses required by the dimension theorem are present. A valid similarity-dimension computation under separation is different from merely solving the equation for an overlapping system.
Dimension Does Not Determine Critical Measure
Proof. For each positive integer \(n\), let \(r_n=\tfrac12-\tfrac1{n+4}\). Then \(0<r_n<1/2\) and \(r_n\to1/2\). Let \(K_n\subseteq[0,1]\) be the attractor of the two similarities \(x\mapsto r_nx\) and \(x\mapsto r_nx+(1-r_n)\). These images are separated because \(r_n<1/2\). By the established dimension formula for separated similarities, \(\dim_H K_n=d_n=\log 2/\log(1/r_n)\). Since \(r_n<1/2\), \(\log(1/r_n)>\log 2\), so \(d_n<1\); since \(r_n\to1/2\), \(d_n\to1\).
At level \(j\), the standard cylinder cover of \(K_n\) has \(2^j\) intervals, each of diameter \(r_n^j\). Its total \(1\)-dimensional cost is \(2^jr_n^j=(2r_n)^j\), which tends to \(0\) because \(2r_n<1\). The diameters also tend to \(0\), so the definition gives \(\mathcal{H}^1(K_n)=0\). Set \(E=\bigcup_{n=1}^\infty K_n\). Countable subadditivity yields \(\mathcal{H}^1(E)\le\sum_n\mathcal{H}^1(K_n)=0\), hence \(\mathcal{H}^1(E)=0\). By countable stability, \(\dim_H E=\sup_n d_n=1\). This proves the claim. \(\square\)
This example rules out the inference “dimension \(1\), therefore positive \(\mathcal{H}^1\).” Dimension records the threshold at which Hausdorff measures change from infinite to zero; it does not specify whether the measure at the threshold is zero, finite and positive, or infinite. For comparison, the interval calculation above gives \(\dim_H[0,1]=1\) and \(\mathcal{H}^1([0,1])=1\). The same dimension can therefore accompany different critical measures.
A Practical Error Check
Before accepting a Hausdorff-dimension calculation, check the following points:
- Is each cover cost computed from diameter, with any radius-to-diameter conversion written explicitly?
- Does the infimum range over all covers admissible at the stated scale?
- Are limits taken in the order required by the definition, with uniform estimates if a bound is meant to hold across scales?
- If a similarity equation is used, have the separation hypotheses of the dimension result been verified?
- Is the conclusion about dimension being kept distinct from a claim about the critical Hausdorff measure?
A particularly common wrong calculation is to cover an interval by \(N\) balls of radius \(1/(2N)\), then use \(N(1/(2N))\) as its \(\mathcal{H}^1\) cost. The correction is to compute each diameter first: \(2(1/(2N))=1/N\), and then sum \(N(1/N)=1\). The arithmetic is short; identifying the quantity in the definition is the essential step.
Check Your Understanding
For each question, state which definition or hypothesis controls the conclusion.
- A ball has radius \(r\). What upper bound does this give for its diameter, and what factor appears when its diameter is raised to the power \(s>0\)?
- In the array \(a_{n,k}\) above, compute \(\lim_n\inf_k a_{n,k}\) and \(\inf_k\lim_n a_{n,k}\) directly from their definitions.
- For the maps \(x\mapsto x/2\) and \(x\mapsto x/3\), why does the similarity equation have a positive root even though the attractor has dimension zero?
- What are \(\dim_H[0,1]\) and \(\mathcal{H}^1([0,1])\) under the diameter-cost convention used here?
- Why does \(\dim_H E=1\) not, by itself, imply \(\mathcal{H}^1(E)>0\)?