Part V
Applications

16 Loeb Measure

Measure theory studies operations that assign magnitudes to sets, like mea suring the length of an interval, the area of a plane region, or the volume of a solid; counting the number of elements in a set; calculating the probabil ity of an event in a sample space or the definite integral of some function over a set; etc. Now, the "measure spaces" on which such operations are defined are typically closed under countable set unions, and this feature is fundamental to the theory. But an internal collection of sets typically fails to be closed in this way. However, in 1973 Peter Loeb discovered that this very failure could be exploited to give a new way of constructing standard measure spaces out of nonstandard entities. 1 This has led to some interesting applications, particularly in probability theory and stochastic analysis. For instance, it provides a representation of Brownian motion as a "random walk with infinitesimal steps"n. We will now develop Loeb's construction, elucidating the role played in it by the nonstandard principles of countable saturation, sequential comprehensiveness, and overflow. We will then apply it to show that Lebesgue measure on the real line can be represented by a weighted counting measure on hyperfinite sets, using infinitesimal weights. But first, a review of some of the basic concepts of measure theory. 1See example 6 of Section 16.1 and example 3 of Section 16.3. 204 16. Loeb Measure 16.1 Rings and Algebras A ring of sets is a nonempty collection A of subsets of a set S that is closed under set differences and unions: • If A, B E A then A -B, A U B E A. It follows that 0 E A, since A-A = 0, and that A is closed under symmetric differences A􀃄B and intersections A n B, since A􀃄B (A-B) U (B(-A), and AnB A-(A-B). An algebra is a ring A that has S EA and hence (indeed equivalently) is closed under complements Ac =S-A. If A is a ring, then AU {S(-A: A E A} is an algebra, the smallest one including A. A a-ring is a ring that is closed under countable unions: • If An E A for all n EN, then UnENAn EA. The equation nneNAn = A1 -(UneN(Al -An)) shows that a a-ring is also closed under countable intersections. A a-algebra is a a-ring that is an algebra. The intersection of any family of a-algebras is a a-algebra. Thus for any A 􀂀 P(S) there is a smallest a-algebra S(A) that includes A. This S(A) is the a-algebra generated by A. Here are some examples of these concepts: (1) P(S) itself is a a-algebra. (2) If S is infinite, then • the collection of all finite subsets of S is a ring that is not an algebra; • the collection of all finite or cofinite subsets of S is an algebra that is not a a-algebra; • the collection of all countable subsets of Sis a a-ring that is not an algebra when S is uncountable. (3) Let CIR be the collection of all subsets of lR that are finite unions of left-open intervals (a, b] = {x E lR : a < x :::; b} with a, b E lR and ao:::; b. (Thus 0 = (a, a] E CJR.) CIR is ring in which each member is in fact a disjoint union of left-open intervals (a, b]. CIR is not an algebra, and is not closed under countable unions: (0, 1) is not in CIR, since each member of CIR will have a greatest element, but (0, 1) is the union of the intervals (0, 1-􀀬] for n EN. 16.1 Rings and Algebras 205 CIR does, however, contain certain significant countable unions: for instance (0, 1] is the union of the pairwise disjoint intervals 􀀭].Any reasonable notion of measure should thus assign to (0, 1] the infinite sum of the measures of the intervals 􀀭]. (4) Let BR be the O"-algebra generated by CR. Each open interval (a, b) in .IRis in BJR, being the union of the countably many left-open intervals (a, b-􀀭] for n EN. Hence every open subset of lR is in BR, being the union of countably many open intervals (take ones with rational end points). On the other hand, if a O"-algebra contains all open intervals, it must contain any left-open (a, b] as the intersection of all (a, b + 􀀭) for n EN. Thus BJR is also the O"-algebra generated by the open intervals, as well as the O"-algebra generated by the open sets of JR. The members of BJR are called the Borel sets. (5) Let S = {1, ... , N} with N an unlimited hypernatural. Then S is hyperfinite, and the collection Pr(S) of all internal subsets of S is an algebra (also hyperfinite) that by transfer of the finite case will be closed under hyperfinite unions, i.e., unions of internal sequences (An : n ::; K) for K E *N. Pr(S) is not, however, a O"-algebra: it contains each initial segment {1, ... , n} with n E N, but does not contain their union because that is the external set N. This same analysis applies to the algebra of internal subsets of any nonstandard hyperfinite set S = {sn : n::; N}. (6) Let A be an algebra in some universe liJ. In any enlargement of lU, *A will be an algebra, by transfer, but in a countably saturated enlargement *A will not in general be a O"-algebra, even if A is. To see this, let (An : n E N) be a sequence of members of *A with union A. Each An is internal, and if A were in A, it would also be internal and hence by countable saturation would be equal to Un k, Amn= 0, since Un:::;kAn and Am are disjoint, and so J.tL(Am) = 0. Hence UnENAn A1 U· · · U Ak, and LnEN J.tL(An) J.tL(AI) +n· · · + J.tL(Ak), from which it follows that Jl.L satisfies Ml. (4) Let A be an internal ring of subsets of some internal set S in a countably saturated enlargement, and let J.tn: A􀇕 *[0, oo] be a finitely additive function. Adapting the construction of (3), put { sh(J.t(A)), if J.t(A) is-limited, Jl.L(A) = oo, if J.t(A) is unlimited or oo. Then reasoning as in (3), we show that Jl.L :A􀜒 [0, oo] is countably additive, and so is a measure on the ring A. (5) This last construction has (3) as a special case, and also covers other natural extensions of (3) that involve hyperfinite summation. Let w: S 􀃷 *JR be an internal "weighting" function on a hyperfinite set S. For each A EPI(S) put (recall the definition of hyperfinite sums in Section 13.19). Then J.tw is a "weighted counting function" that is finitely additive and induces the measure J.t£ on PI(S). In fact, every internal finitely additive function J.t : PI(S) 􀃷 *[0, oo] arises in this way: put w(s) = J.t({s}). Example (3) itself is the special case of a uniform weighting in which each point is assigned the same weight w(s) = 16.3 Outer Measures We now review the classical procedure of Caratheodory for extending a measure J.t on a ring of sets A to a measure on a a-algebra including A. If B is an arbitrary subset of the set S on which A is based, put Here the infimum is taken over all sequences (An : n E N) of elements of Athat cover B. The function J.t+ : P(S) 􀀆 [0, oo] is called the outer measure defined by J.t (although it may not actually be a measure). It has the following properties: 16.3 Outer Measures 209 • J.L+ agrees with J.L on A: if B E A, then J.L+ (B) = J.L( B). In particular, J.L+(0) = 0. • Monotonicity: if A􀁕 B, then J.L+(A) 􀀢 J.L+(B). • Countable subadditivity: for any sequence (An) of subsets of S, • For any B 􀁕 S and any c E JR+ there is an increasing sequence A1 􀁽 A2 􀁕 · · · of A-elements that covers Bnand has A set B 􀁕 S is called J.L+ -measurable if it splits every set E C S J.L+additively, in the sense that For this to hold it is enough that whenever J.L+(E) < oo. The class A(J.L) of all J.L+ -measurable sets has the following properties. • A(J.L) is a a-algebra. • A 􀁕 A(J.L), i.e., all members of A are J.L+ -measurable. Hence A(J.L) includes the a-algebra S(A) generated by A. • All J.L +-null sets belong to A(J.L) . • J.L+ is a measure on A(J.L), and hence is a measure on S(A). • If J.L is a-finite on A, and A is an algebra, then J.L+ is the only extension of J.L to a measure on S(A) or on A(J.L). Because A(J.L) contains all J.L+ -null sets, J.L+ is a complete measure on it, which means that • if A􀁕 B E A(J.L) and J.L+(B) = 0, then A E A(J.L). This entails that • if A, BE A(J.L) with A􀁕 Bnand J.L+(A) = J.L+(B), then any subset of B-A belongs to A(J.L) (and is J.L+-null), and hence any set C with A􀁕 C 􀁕 B belongs to A(J.L) and has J.L+(c) = J.L+(A) = J.L+(B). 210 16. Loeb Measure 16.4 Lebesgue Measure Lebesgue measure is defined by the outer measure ..\ + constructed from the measure A on CIR that is determined by putting ..\( (a, b]) = b -a. The members of the a-algebra CJR(..\) are known as the Lebesgue measurable sets and include all members of the a-algebra BIR of Borel sets generated by CJR. We will write ..\(B) for ..\+(B) whenever B is Lebesgue measurable. Some facts about Lebesgue measure that will be needed are: (1) A is the only measure on BIR that has ..\((a, b))U= b-a: any measure on an algebra including BIR that agrees with A on open intervals must agree with ..\ on all Borel sets. (2) For any Lebesgue measurable set B there exist Borel sets C, D with C 􀃙 B 􀁕 D and ..\(Dn-C) = 0, hence ..\(B)n= ..\(C) =n..\(D). (3) A set B 􀁕 IR is Lebesgue measurable if for each c E JR+ there is a closed set Cc 􀃙Band an open set De :2 B such that ..\(Den-Cc) ..(B)n= JLL(S n *B) holds in general, but that suggestion is quickly dispelled by the case B = Q. Every grid point is a hyperrational number, so S 􀁽 *Q and hence JLL(S n *Q) = JLL(S) = oo, while >..(Q) = 0. Rather than S n * B, the appropriate set to represent B in S is the set of grid points that approximate members of B infinitely closely. This is the set {s E S : s is infinitely close to some r E B} {s E S: sis limited and sh(s) E B}, which may be called the inverse shadow of B. The definition of sh -l(B) uses a condition that is not internal, so the set itself cannot be guaranteed to be internal, and more strongly may not be Loeb measurable, i.e., may not belong to A(JLL). One case in which it is not internal but nonetheless is Loeb measurable occurs when B = 􀁴: since sh-1(R) = {s E S: sis limited}n= UnEN(S n *(-n,nn)), while each set Sn*( -n, n) is an internal subset of Sand so belongs to A, it follows that sh-1(R) belongs to A(JLL) by closure under countable unions. But sh-\R) cannot be internal, because it is bounded in *􀁴 but has no least upper (or greatest lower) bound. The general situation is this: 16.8 Lebesgue Measure via Loeb Measure 217 Theorem 16.8.2 A subset B of JR is Lebesgue measurable if and only if sh-1 (B) is Loeb measurable. When this holds, the Lebesgue measure of B is equal to the Loeb measure of the set of grid points infinitely close to points of B: >.(B) = J.LL (sh-1 (B)). Proof Let M = {B 􀃚 JR: sh-1(B) E A(J..t£)}. FornB EM, put v(B) = J..lL (sh-1(B)). Our task is to show that M is the class CJR(>.) of Lebesgue measurable sets, and that v is the Lebesgue measure >.. By properties of inverse images of functions, sh-1 (0) -0, sh-1(A-B) sh-1(A) -sh-1 (B), sh-1(UnENAn) UnENsh-l(An) · Since A(J.LL ) contains 0 and is closed under set differences and countable unions, these facts imply that M has the same closure properties. Since sh-1n(1R) E A(J.LL ), as was shown above, we also have JR E M. Altogether then M is a a-algebra, on which v proves to be a measure. At this point we need the the following lemma. Lemma 16.8.3 M includes the Borel algebra BJR, and v agrees with Lebesgue measure on all Borel sets. Proof Each open interval (a, b) 􀃮 JR belongs to M, since sh-l ( (a, b)) is the union of the sequence (An : n EnN), where A = S n *(a + .! b -.! ) E A. n . n' n But BJR is the smallest a-algebra containing all open intervals (a, b), so this implies that BR 􀃚 M. Also, by Theorem 16.8.1, J.LL(An) = (b-􀈅)-(a+􀛝) = b-a-􀛞 ' and hence as the An 's form an increasing sequence, v((a, b))o= j..tL (sh-1 ((a, b))o= lim j..tL(An) = b-a. n􀁝oo Thus v is a measure on BJR that agrees with A on all open intervals. But any such measure must agree with >.non all Borel sets (16.4(1)). D Now, if B 􀂀 JR is Lebesgue measurable, then (16.4(2)) there are Borel sets C, D with C 􀃚 B 􀃮 D and >.(C)o= >.(B)o= >.(D). Then sh-1n(C) 􀃚 sh-1n(B) 􀂀 sh-1 (D), 218 16. Loeb Measure and by Lemma 16.8.3, C,D EM, whence sh-1(C),sh-1(D) E A(JLL), and JLL(sh-1(C)) = v(C) =n>.(C)n= >.(D) = v(D) = JLL(sh-1n(D)). Since J.lL is a complete measure on A(JLL) (by the general theory of outer measures), it follows that sh-1(B) E A(JLL), and hence BE M, with 1n v(B) = JLL(sh-1(B)) = JLL(sh-(C)) = .A(C) =n>.(B). This establishes that every Lebesgue measurable set is in M (i.e., CIR(.A) 􀃮 M) and that v agrees with >. on all Lebesgue measurable sets. It remains now to show that M 􀃮 CIR(.A), and for this we need the result from Section 11.13 that the shadow of any internal subset of *IR is topologically closed as a subset of IR, and so is a Borel set. Let B E M, i.e., sh-1(B) E A(JLL)· First we consider the case that sh-1(B) is JLL-finite and show that B is Lebesgue measurable by the criterion 16.4(3). But if JLL(sh-1(B)) < oo, then by Lemma 16.6.1, sh-1(B) is JL-approximable, so for any given c E JR+ there exist sets C, D E A with C 􀃮 sh-1(B) 􀃮 D and JLL(D)n-JLL(C) ..(B) of a subset B of lR can be obtained as the Loeb measure JLL(sh-1(B)). D