Measure Theory
The Vitali Covering Lemma: Extracting Disjoint Balls
Hand someone a wild, infinite pile of overlapping balls covering a set in ℝⁿ — and the Vitali Covering Lemma lets them pull out a countable, pairwise-disjoint subfamily that, once each ball is inflated by a fixed factor of 5, still covers everything. That single trick — trade a factor of 5 in radius for perfect disjointness — is the combinatorial engine behind the Hardy–Littlewood maximal inequality, the Lebesgue differentiation theorem, and much of geometric measure theory.
Precisely: from any collection of balls with uniformly bounded radii, there is a countable disjoint subcollection {Bⱼ} such that the union of the original balls is contained in ⋃ⱼ 5Bⱼ, where 5Bⱼ is Bⱼ dilated about its center by 5. The point is that disjointness is free — you only pay a bounded, dimension-independent enlargement.
- FieldMeasure theory / real analysis
- Named afterGiuseppe Vitali (1908)
- Key hypothesisUniformly bounded radii (sup of radii finite)
- StatementDisjoint subfamily {Bⱼ} with ⋃B ⊆ ⋃ 5Bⱼ
- Proof techniqueGreedy selection by radius (Zorn / transfinite or maximal disjoint family)
- Generalizes toMetric spaces; Besicovitch & Vitali covering theorems
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
The precise statement
Let 𝓕 be any collection (finite, countable, or uncountable) of open balls in ℝⁿ — or in a general metric space — with uniformly bounded radii: R := sup{ radius(B) : B ∈ 𝓕 } < ∞. Then there exists a countable, pairwise-disjoint subcollection 𝓖 = {B₁, B₂, …} ⊆ 𝓕 such that
- every ball B ∈ 𝓕 meets some Bⱼ ∈ 𝓖 with radius(Bⱼ) ≥ ½ radius(B), and
- ⋃_{B ∈ 𝓕} B ⊆ ⋃ⱼ 5Bⱼ, where 5Bⱼ denotes the ball concentric with Bⱼ but with five times the radius.
The number 5 is not sacred: any constant c > 3 works if you select radii within a factor arbitrarily close to their supremum. With a strict greedy choice (each selected ball at least half the sup of remaining radii) the clean constant is 5. Note there is no measure hypothesis here — this is a purely geometric/combinatorial statement.
The picture: why a factor of 5 buys disjointness
Imagine a chaotic heap of disks blanketing a region, overlapping every which way. You want a tidy, non-overlapping sub-selection, but of course a disjoint subfamily can't literally cover the same area — gaps appear between the chosen disks. The lemma's insight is that those gaps are controlled: any disk you threw away was blocked by an already-chosen disk that is at least half as big and touches it.
Now do the geometry. If a discarded ball B (radius r) meets a kept ball Bⱼ (radius rⱼ ≥ r/2), then every point of B lies within distance 2r + rⱼ of Bⱼ's center. Since 2r ≤ 4rⱼ, that distance is ≤ 5rⱼ. So B ⊆ 5Bⱼ. The factor 5 is exactly 4 (from the discarded ball being at most twice the radius) plus 1 (the kept ball itself). Disjointness costs you a fivefold inflation and nothing more — crucially, independent of dimension n.
The key idea of the proof
The mechanism is greedy selection, largest-first. Partition 𝓕 by radius into shells 𝓕ₖ = { B : R/2ᵏ⁺¹ < radius(B) ≤ R/2ᵏ } for k = 0, 1, 2, …. Build 𝓖 in stages:
- Stage 0: choose a maximal pairwise-disjoint subfamily 𝓖₀ of 𝓕₀ (exists by Zorn's lemma).
- Stage k: from 𝓕ₖ, choose a maximal disjoint subfamily 𝓖ₖ among the balls that are also disjoint from everything already chosen in 𝓖₀ ∪ … ∪ 𝓖ₖ₋₁.
Set 𝓖 = ⋃ₖ 𝓖ₖ; it is disjoint by construction. Countability follows because in ℝⁿ disjoint balls of radius > ε have centers separated by > ε, so only countably many fit. Covering claim: take any B ∈ 𝓕, say B ∈ 𝓕ₖ. By maximality of 𝓖ₖ, B must intersect some already-chosen ball Bⱼ of radius > R/2ᵏ⁺¹ ≥ ½ radius(B). By the distance estimate above, B ⊆ 5Bⱼ. That's the whole argument.
A worked special case
Take 𝓕 to be the intervals (balls in ℝ¹) Iₐ = (a − 1, a + 1) for every a ∈ [0, 10] — an uncountable open cover of the open interval (−1, 11) by overlapping length-2 intervals, all radius 1. Greedy selection (all radii equal, so any maximal disjoint choice works) picks, say, I₀ = (−1, 1), then I₂ = (1, 3), then I₄, I₆, I₈, I₁₀ — six disjoint intervals.
These six do not cover (−1, 11): the point 1 sits in a gap. But the 5-dilations 5I₀ = (−5, 5), 5I₂ = (−3, 7), … each stretch to radius 5 and together swallow the whole union and far more. From uncountably many intervals we extracted a disjoint family of six, at the price of a factor-5 fattening. This is exactly the trade the maximal-function proof exploits: replace an unmanageable cover by finitely/countably many disjoint balls whose measures you can simply add up.
Why the hypotheses matter
Bounded radii is essential. Drop it and greedy-largest-first has no largest to start with. Counterexample: in ℝ¹ take the family { (−r, r) : r > 0 }. Any two of these intersect (they all contain 0), so a disjoint subfamily has at most one member — a single ball — and no finite dilation of one ball covers ⋃ (−r, r) = ℝ. So without sup radius < ∞ the conclusion fails outright.
Distinguish this from the Vitali covering theorem. The 5r-lemma keeps a factor of 5; the covering theorem instead assumes a fine (Vitali) cover — every point is covered by balls of arbitrarily small radius — and a set of finite outer measure, and then extracts a disjoint subfamily that covers the set up to a null set with no dilation at all. Both are proved by the same greedy idea; the fine-cover hypothesis is what lets you iterate and drive the leftover measure to zero. The Besicovitch covering theorem is a different beast: it keeps radius 1 but pays with bounded overlap Cₙ instead of disjointness, and needs balls centered on the set.
Applications and significance
The 5r-lemma is the disjointification workhorse of analysis. Its headline consequence is the Hardy–Littlewood maximal inequality: |{ Mf > λ }| ≤ (5ⁿ/λ) ‖f‖₁. The proof covers the level set by balls on which the average of |f| exceeds λ, extracts a disjoint subfamily via Vitali, and sums their measures — the 5ⁿ is the dilation factor raised to the dimension.
- Lebesgue differentiation theorem: the maximal inequality gives that averages of an L¹ function over shrinking balls converge to the function a.e.
- Geometric measure theory: density theorems, rectifiability, and Hausdorff-measure comparisons all lean on Vitali-type covers.
- Calderón–Zygmund theory & harmonic analysis: the maximal function it delivers underlies singular-integral bounds.
Because the constant is dimension-independent (the 5 doesn't grow with n; only the volume factor 5ⁿ does), the lemma scales cleanly and even survives in doubling metric-measure spaces — which is why it, not Besicovitch, is the default tool in modern analysis on metric spaces.
| Result | What you get | Key hypothesis | Enlargement factor |
|---|---|---|---|
| Basic Vitali (5r) covering lemma | Countable disjoint subfamily; 5-dilations cover the union | Radii uniformly bounded | 5 (any c > 3 works) |
| Vitali covering theorem | Disjoint subfamily covers the set up to measure zero — no dilation | Fine (Vitali) cover; finite outer measure | 1 (none) |
| Besicovitch covering theorem | Cover of centers by boundedly-many disjoint subfamilies | Balls centered at points of the set, radii bounded | 1 (bounded overlap Cₙ) |
| Hardy–Littlewood application | Weak (1,1) bound for the maximal function | Uses the 5r lemma directly | 5ⁿ appears in the constant |
Frequently asked questions
Why the factor 5 specifically — and can it be smaller?
The 5 comes from 4 + 1: a discarded ball is at most twice the radius of the kept ball it hits (contributing 2r ≤ 4rⱼ to the distance) plus the kept ball's own radius rⱼ. Any constant c > 3 works if you select balls whose radii are within a factor approaching the supremum of the remaining radii; the clean 5 corresponds to the 'at least half the sup' greedy rule. You cannot push it down to 1 — that would require the disjoint balls themselves to cover, which is impossible.
Is the extracted subfamily always countable, even if 𝓕 is uncountable?
Yes, in ℝⁿ (and any separable or doubling metric space). Pairwise-disjoint balls of radius greater than ε have centers separated by more than ε, and only countably many such centers fit in a separable space. So each radius-shell contributes countably many balls, and the union over countably many shells stays countable. This is why the lemma yields a countable disjoint family from an arbitrary collection.
What breaks if the radii are not uniformly bounded?
The greedy 'largest-first' selection has no largest ball to begin with, and the conclusion genuinely fails. The family { (−r, r) : r > 0 } in ℝ¹ is the textbook counterexample: every pair intersects (all contain 0), so any disjoint subfamily is a single interval, and no fixed dilation of one bounded interval can cover the union ℝ. Bounded radii is therefore an essential hypothesis, not a technical convenience.
How is this different from the Vitali covering theorem?
The 5r-lemma is purely geometric: it keeps a dilation factor of 5 and needs only bounded radii, with no measure assumption. The covering theorem is measure-theoretic: it assumes a fine (Vitali) cover — every point covered by balls of arbitrarily small radius — and finite outer measure, then produces a disjoint subfamily covering the set up to a null set with no dilation. The theorem is the sharper, measure-consuming refinement built on the lemma's engine.
Does the lemma hold in general metric spaces or infinite dimensions?
The 5r covering lemma holds in any metric space, since its proof is purely combinatorial-geometric and never uses a measure. Countability of the extracted family, however, needs separability (or a doubling condition). Where things get subtle is the Besicovitch covering theorem, which can fail in infinite-dimensional spaces and in some metrics; the Vitali 5r-lemma's robustness is exactly why it, not Besicovitch, is the standard tool for analysis on general metric-measure spaces.
Where does the 5ⁿ in the maximal inequality come from?
In proving the weak (1,1) bound |{Mf > λ}| ≤ (5ⁿ/λ)‖f‖₁, you cover the level set by balls whose average of |f| exceeds λ, then apply the 5r-lemma to get a disjoint subfamily {Bⱼ}. Since the original union sits inside ⋃ 5Bⱼ, and |5Bⱼ| = 5ⁿ|Bⱼ| in ℝⁿ, the dilation factor 5 raised to the dimension n multiplies the summed measures of the disjoint balls. Disjointness lets you add |Bⱼ| ≤ (1/λ)∫_{Bⱼ}|f| without double-counting.