A measure may assign zero to a set without its -algebra containing every subset of that set. Completing the measure adds exactly those missing subsets, without changing the values of previously measurable sets. For Lebesgue measure this creates measurable sets that are not Borel. Even completion does not make every subset of the line measurable: a different construction gives a set whose inclusion in a hypothetical measure on all subsets would conflict with translation invariance and countable additivity while extending Lebesgue measure.
Completing a measure
Let be a measure space.
Definition 1.9.1
(Complete measure and completion)
The measure is complete if every subset of every measurable null set is measurable. The completion of is the pair , where
and .
Here need not belong to . Equivalently, exactly when is contained in a measurable null set for some : the reverse direction follows by writing .
Theorem 1.9.2
(Completion theorem)
is a -algebra and is a well-defined complete measure extending . It is the unique complete extension of on this -algebra, and every complete extension of has a domain containing and agrees with there.
Proof
(Well-definedness, closure, additivity and minimality)
If , where and , then the measurable set lies in and lies in . Both have measure zero. Decomposing into its intersection with the other set and its difference shows , even if the common value is infinite. Thus the definition is unambiguous.
For complements, if with null, then
where the first set is measurable and the second is a subset of . Countable unions have the same form because the union of countably many measurable null sets is measurable and null. Hence is a -algebra.
If are disjoint completed sets, the measurable are disjoint. Their union represents up to a subset of the null set , so countable additivity follows from that of . Taking shows extension. If , then lies inside the measurable null set ; every subset of is therefore in and has measure zero. Thus the extension is complete.
Finally, in any complete extension on a larger -algebra, every is measurable and null, so every lies in its domain. It agrees with on ; for , monotonicity forces its value on to be . This proves both minimality of the domain and uniqueness on it.
Proof X-ray
Why every new value is forced
Every step uses one fact: a completed set is sandwiched as with . Well-definedness compares two sandwiches by intersection rather than subtraction, so it survives . Closure under complements needs the explicit formula for , because the complement of a non-measurable is not available directly. Additivity uses countable unions of null sets. The key move for uniqueness is monotonicity applied to the sandwich: any complete extension must give a value between and . No value is chosen; each is forced.
Common mistake
(Subtracting infinite measures)
To show it is tempting to write . If this expression is meaningless. Compare both with instead, as the proof does.
Worked example 1.9.3
(Completing a measure on three points)
Let , and , . This measure is not complete: is not in . The measurable null sets are and , so ranges over all subsets of , and is the power set of . For example gives , and .
Worked example 1.9.4
(Completing a point mass on the line)
Let be the Dirac measure at on the Borel sets of . The Borel set is null. Every equals , a Borel set together with a subset of . So the completion is defined on every subset of , with if and otherwise. In this edge case completion produces the whole power set.
Borel and Lebesgue sets
Theorem 1.9.5
(Lebesgue measure completes Borel measure)
Lebesgue measure on is the completion of its restriction to the Borel -algebra.
Proof
(Minimality one way, regularity the other)
Lebesgue measure is a complete extension of the Borel restriction, so minimality puts the completion inside the Lebesgue -algebra. Conversely, use regularity: for finite-measure Lebesgue , choose compact with . Their union is Borel, , and is null. By outer regularity choose open with ; the Borel set contains and has measure zero. Hence is in the completion. For arbitrary , first split into bounded pieces, apply the same argument to each, and take countable unions of the Borel pieces and Borel null covers.
The inclusion of Borel sets in Lebesgue measurable sets is strict, and the witness lives inside a null set that is nevertheless large.
Counterexample mode
A null set need not be countable
Claim: a Lebesgue null set must be countable. Let be the middle-thirds Cantor set. It is compact, and at stage it lies in closed intervals of total length ; hence . For every binary sequence , the point lies in , and two different sequences give different points: at the first differing digit the difference is , while the remaining digits change the value by at most . So contains an injective image of and is uncountable. The true statement goes one way only: countable sets are null.
Worked example 1.9.6
(A Cantor point that is not an endpoint)
The removed-interval endpoints of have finite ternary expansions, but contains more. Take . Then , so . Its ternary expansion never terminates, so is not an endpoint of any removed interval. The digit map therefore reaches points that the construction never names directly.
Conversely, each point of belongs to a nested sequence of retained left/right intervals, whose diameters tend to zero, yielding such a digit expansion. Combined with the injective digit map above, has exactly the cardinality .
Theorem 1.9.7
(Lebesgue measurable sets that are not Borel)
In ZFC, some subset of the Cantor set is not Borel. Every such subset is Lebesgue measurable with measure zero.
Proof
(Counting Borel sets)
We work in standard ZFC, including the countable choices used in coding. There are at most Borel subsets of . One standard counting argument codes Borel operations by countable well-founded trees: leaves carry rational-endpoint interval labels, while internal nodes represent complement or countable union. Such codes are countable labeled objects; their number is at most , and they can be encoded by subsets of ; the coded sets form a -algebra containing the open sets. Cantor's theorem gives . Consequently some subset of is not Borel. It is nevertheless Lebesgue measurable with measure zero by completeness. This is an existence argument, not an explicit description of a particular non-Borel set.
A set outside the completion
Theorem 1.9.8
(A Vitali set is not Lebesgue measurable)
Using the axiom of choice, there is a set that is not Lebesgue measurable.
Proof
(Rational translates tile an interval)
On define when . Choose one representative from each equivalence class and call the resulting set . For distinct rationals , the translates and are disjoint: an intersection would give two representatives in the same class.
Every has a representative with . Hence
Suppose were Lebesgue measurable. Translation invariance would give the same finite measure to each translate. If , countable additivity makes the middle union null, contradicting its containment of . If , the disjoint countably infinite union has infinite measure, contradicting its containment in . Thus is not Lebesgue measurable and is not in the completed Borel -algebra.
Common mistake
(Expecting completion to make every set measurable)
Completion adds only subsets of null sets. The Vitali set is not contained in any null set with a measurable remainder, so it stays outside. The two examples make different points: a non-Borel subset of is measurable because it lies in a null set, whereas cannot be measurable at all. Neither conclusion follows merely from the other.
Practice
Quick check
(If , and , what is ?)
Solution · Answer
Zero, since is contained in the measurable null set .
Quick check
(Why does completion not alter the measure of an old set ?)
Solution · Answer
Represent ; well-definedness makes every other representation give the same value .
Quick check
(Why can every subset of the Cantor set be Lebesgue measurable even though not every such subset is Borel?)
Solution · Answer
The Cantor set is a Lebesgue null set and Lebesgue measure is complete; the cardinality argument shows the Borel family is too small to contain all its subsets.
Quick check
(In the Vitali argument, why are the rational translates pairwise disjoint?)
Solution · Answer
If , then , so the representative property forces and then .