Mathematical Analysis II - Part B
2 Integral calculus in several variables
In this section we approach the problem of integrating real functions of several real varia- bles. The knowledge of Riemann integral for functions of one real variable is required.
2.1 Lebesgue versus Riemann integral: motivation and a brief presentation
A theory of Riemann integration for functions of several real variables (i.e. f : A R, ! with A Rn) could be constructed by adapting in a natural way the already known theory ⇢ for one variable: the role of intervals of R will be taken by the “n-dimensional intervals” I = I1 In,wheretheIj are intervals of R with extremities included or not; there ⇥···⇥ will still be the simple functions (or “step functions”, “staircase functions”) from which to start the definition of integral, the upper and lower sums, and so on.
Figure 2.1: The integral of a function of two real variables.
The problem of defining the integral is substantially equivalent to the one of providing a measure to subsets of Rn: one would say that E Rn is elementarily measurabile ⇢ (or measurable `ala Peano-Jordan) when its indicator function E is Riemann-integrable, calling “measure of E” (or also “area”, “volume” in the cases n =2, 3) the value of such integral. Obvious good qualities of this theory are its simplicity and naturality; however, the drawbacks already observed in the case of one variable are still present. Let us mention three of them. Riemann-integrable functions (and, in parallel, elementarily measurable subsets) are • relatively few when one has to enlarge its scope beyond continuous functions, or rat- her beyond the functions having a “reasonable number of singular points”.(13) Na-
(13)This is what happens since more than one century in the progress of exact sciences, as new important
Corrado Marastoni 29 Mathematical Analysis II - Part B
mely, on a compact n-dimensional interval a Riemann-integrable function must be bounded, and continuous “almost everywhere”:(14) for example, this fact says that the Dirichlet function (which is discontinuous in all points [0, 1]) cannot be Q [0,1] Riemann-integrable. \ The Riemann integral has no satisfactory interaction with the limit.Givenasequence • of integrable functions fn which converges to a function f in a su ciently weak way (e.g. pointwise, or even less), we would like also f to be integrable and that it could be possible to “pass to limit under the sign of integral”, i.e. that f =lim fn. Now, in such a generality, this request is rather optimistic, because a weak convergence cannot R R control “mass dispersion” phenomena in the domain, as the following example shows.
n n (n 1) Example. Let us define fn : R R as 2 on the interval [2 , 2 ], and 0 elsewhere. Any of ! the fn is Riemann-integrable with fn = 1, and the sequence (fn) converges pointwise to the zero
function f 0, which is obviously Riemann-integrable as well; but f =0=lim fn =1. ⌘ R 6 R R To prevent such phenomena, a reasonable request is that this should be true at least for sequences “dominated” by a fixed integrable function, i.e. such that there exists an integrable function g such that fn g for any n N. However, for the Riemann | | 2 integral also this is not enough: in fact the convergence should be uniform, because pointwise convergence could not work.
Example. Let q : N Q be any bijection (we know that Q is countable), and define fn : R R ! ! as 1 onlu on the rationals qm with m n, and 0 elsewhere: any of the fn is Riemann-integrable (it has finitely many discontinuity points) with fn =0,butthesequence(fn) converges pointwise (not
uniformly!) to Dirichlet function f = [0,1] , which is not Riemann-integrable. Q\ R
The Riemann integral is linked to the properties of real numbers. Riemann’s costruction • depends a lot on the properties of real numbers (in particular on the total order), and the extendibility to the case of real functions defined on more general sets than Rn appears to be di cult.
These reasons, among some others, suggested at the end of XIXth century to look for a more e↵ective integration theory, being prepared to pay a price in terms of simplicity but without loosing too much of its naturality: among various attempts, the one by Henri tools (for example, Fourier series) could generate highly irregular functions. (14)The precise meaning of this “almost everywhere” is —as we shall see later— “everywhere excepted a null set”, where “null set” means a set of points whose Lebesgue measure is zero: the exact statement is n that a bounded function with compact support f : R R is Riemann-integrable if and only if the set of its ! discontinuity points is a null set. The Lebesgue measure, as we shall see, is much larger than the elementary measure of Peano-Jordan. For example, all countable sets are Lebesgue-measurable (with measure zero), 1 while they could be not elementarily measurable: for example E = n : n N is elementarily measurable 1 1 { 12 } (if sn : R R is the simple function with values 1 on [0, n ] n 1 ,..., 2 , 1 and zero elsewhere, it holds ! 1 [{ } E sn and hence sn = is a upper sum for E ), while the Dirichlet set Q [0, 1] is not. A classical R n \ example of Riemann-integrable function whose discontinuity set is Lebesgue-null but not elementarily R p measurable is f :[0, 1] R defined as 0 on [0, 1] Q and as 1 on Q [0, 1] with p, q 0 coprime, ! \ q q 2 \ whose discontinuity set is the Dirichlet set Q [0, 1]. \
Corrado Marastoni 30 Mathematical Analysis II - Part B
Lebesgue’s (1875-1941) eventually turned out to be the most successful. Let us discuss why by making reference to the just mentioned Riemann’s drawbacks. When compared to Riemann theory, Lebesgue’s considerably enlarges the family of • measurable sets (and, in parallel, the family of measurable functions, possibly integra- ble), so that it is very complicated to describe non Lebesgue-measurable subsets of Rn; in particular, the Dirichlet function is Lebesgue-integrable with integral zero. Q [0,1] The role played by continuity in Riemann\ integral is played here by the property of “measurability” which is extremely weak and in practice satified all the time. Lebes- gue’s construction, which nowadays keeps having a large agreement(15), is in a certain sense the result of a process of completion of Riemann’s, similar to the construction of real numbers starting from rationals. The Lebesgue integral works much better under limit with respect to Riemann’s: if • we have a sequence of Lebesgue-integrable functions fn which converges pointwise “almost everywhere”(16) to a function f, in two very frequent cases (a monotone or a dominated sequence) also f is Lebesgue-integrable and f =lim fn. In particular
the example with f = [0,1] is fixed (the sequence fn converges monotonically to f, Q\ R R which is Lebesgue-integrable and 0 = =lim f ), while the example with Q [0,1] n the “mass dispersion” keeps resisting (actually\ the sequence f converges pointwise R R n to f = 0 but neither in a monotone nor in a dominated way). Lebesgue theory is largely independent from particular properties (topological or not) • of the domain, and hence it naturally suits to real functions defined on any domain: in fact, to present the theory in full generality does not create more di culties with respect to Rn (on the contrary, it rather helps to better clarify the situation). Anyway, we should point out that a large part of the results which are valid for • Lebesgue theory are valid also for Riemann’s: but usually the hypotheses in the latter are some heavier (and, after all, quite unessential).
For what has been said, it should appear convenient to make reference to Lebesgue integral from now on. Let us briefly recall the main aspects of this theory, without proofs.
Let X be any set, and let us plan to introduce a measure µ of the subsets of X: in other words, to a A X we want to associate a measure µ(A), that we expect to be a number ⇢ 0, possibly + . How can we do? The first requirement would be, at one side, to be able 1 to measure as many subsets of X as possible; but, on the other side, the measure should satisfy some natural properties, that we now list.
(15)We should however say that, while Lebesgue theory fixes many problems, some unexpected others sin x appear: for example x is not Lebesgue-integrable on R although it has a generalized Riemann-integral (called “Dirichlet integral”); in fact it is not absolutely Riemann-integrable. (16)i.e., as said above, “excepted a null set”: it is clearly a rather weak notion of convergence.
Corrado Marastoni 31 Mathematical Analysis II - Part B
– Enough stability for the set-theoretical operations . We aim that the family of µ- measurable subsets, beyond than being rich of elements, would be stable under the main set-theoretical operations, in particular under the union and the intersection (even better, under countable union and intersection) and under complementation in X. In other words: if A, B X are µ-measurable, we aim that also A B, A B and ⇢ [ \ {X A = X A be so; and that, if An is a countable family of µ-measurable subsets, \ (17) also An and An be so. – Isotony .IfA B, it would be natural that µ(A) µ(B). T ⇢S – Additivity . The measure should be additive on disjoint unions, better on countable ones. In other words, if A, B X are µ-measurable and disjoint, it should hold that ⇢ µ(A B)=µ(A)+µ(B); moreover, even better, if A is a countable family of µ- [ n measurable pairwise disjoint subsets, then µ( An)= µ(An). S P In fact, Lebesgue theory is developed in full generality (and without problems, rather with a better clarity and language) in the framework of a measure space (X, ,µ), where X Measure space M is any set, a family of parts of X which contains the same X and is stable under com- M plementation and countable union,(18) and µ : [0, + ] is a function non identically M ! 1 equal to + and countably additive, i.e. such that given a countable family An of 1 (19) 2M pairwise disjoint subsets of X it holds µ( An)= µ(An). We shall say that µ is a (positive) measure on X, and the elements of (called a -algebra of parts of X)willbe Measure S M P called the µ-measurable subsets of X. Measurable sets
Examples. (0) (Measure zero) The first (trivial) example is the measure identically 0. (1) (Measure which counts the elements) Choose = P (X), and for A X define µ(A)asthenumberofelementsofA (set M ⇢ µ(A)=+ if A is not finite). (2) (Measure concentrated in a point) Choose = P (X); fixed an element 1 M x0 X, for A X define µ(A)=0ifx0 / A and µ(A)=1ifx0 A. 2 ⇢ 2 2
From now on let us concentrate on X = Rn, with the aim of constructing the Lebesgue n measure n on it, without forgetting that on R we already have Riemann’s elementary measure to be preserved as far as possible. Hence, in this particular framework, to the previous desired properties we should also add the following familiar ones. – Extension of elementary measure . We aim the elementarily measurable subsets of Rn be also n-measurable, and that the two measures coincide on them. In particular, n for a bounded interval I = j=1 ]aj,bj[ (with or without extremities) it should be n n(I)= (bj aj). j=1 Q – Invariance under isometries . The measure of A Rn should not change by applying Q ⇢ (17)The measurability of countable unions would be a first evident progress with respect to Riemann: recall that the single points are elementarily measurable (with measure zero), but their countable union Q [0, 1] (the Dirichlet set) is not. \ (18)As a consequence, there is stability also under countable intersection (by de Morgan duality) and obviously also under finite intersection; but also under di↵erence, because A B = A (X B). (19) \ \ \ It is then easy to show that one also has µ(?)=0,andthatalsothefiniteadditivity(i.e.ifA1,...,An n 2 are pairwise disjoint then µ(A1 An)= µ(Aj )) and isotony (i.e. if A, B with A B M [···[ j=1 2M ⇢ then µ(A) µ(B)). P
Corrado Marastoni 32 Mathematical Analysis II - Part B
to A any isometry of Rn, i.e. a function of Rn into itself which preserves the distances (e.g. the translation by a fixed vector).
Let us start the construction of n (in particular, of the -algebra on which it is defined). n The first step is to define n on the intervals: if I = j=1 ]aj,bj[ (possibly half open or n also closed; and possibly degenerated), we set n(I):= (bj aj) , as expected. The Q j=1 second is, then, to define a quantity (called external measure) which makes sense for External measure n⇤ Q any subset: given A Rn,weset ⇢ + + 1 n 1 n⇤ (A):=inf n(Ik): (Ik)k a sequence of intervals of R such that A Ik . R 2N ⇢ ⇢ k=0 k=0 P S In other words,e we consider the various at most countable covers of A made by n-dimen- sional intervals, compute the sum of the measures of the intervals of the cover, and then take the infimum of such sums among the various covers.(20) n Proposition 2.1.1. The external measure ⇤ : P (R ) [0, + ] : n ! 1 (a) respects the measure of the intervals (i.e. if I is an interval then n⇤ (I)= n(I));
(b) it holds n⇤ (?)=0; (c) is isotone (i.e. if A B then (A) (B)); ⇢ n⇤ n⇤ (d) is countably subadditive (i.e.: given (Ak)k it holds n⇤ ( Ak) n⇤ (Ak)). 2N Unfortunately the external measure n⇤ is not countably additive,S becauseP there exist (but we shall not prove it) some very technical examples of sequences (Ak) of pairwise disjoint subsets for which n⇤ ( Ak) < n⇤ (Ak). However, since it is unreasonable to give up this n property, the idea is to restrict ⇤ to a selected family of subsets of R where everything S P n works well. It turns out that this selected family is very large and satisfactory, as the following result shows. n Theorem - Definition 2.1.2. (Carath´eodory) Let n be the family of subsets of R M which “decompose additively the external measure”:
n n n = A R : ⇤ (E)= ⇤ (E A)+ ⇤ (E A) for any E R . M { ⇢ n n \ n \ ⇢ } Then: (a) is a -algebra containing all intervals, all subsets having zero external measure Mn (among them, the finite or countable ones), and in general all open and all closed euclidean subsets of Rn;
(b) the restriction of n⇤ to n is a measure, in particular is countably additive (i.e.: M n given in n a countable family (Ak)k of pairwise disjoint subsets of R it holds M 2N n⇤ ( Ak)= n⇤ (Ak)). Such measure extends Peano-Jordan’s, and is invariant under isometries (in particular under translations).(21) S P Lebesgue- measurable (20)Note that it is not necessary to require that the covers should be formed by pairwise disjoint intervals (we could also require it: in any case, after taking the infimum, the result would not change). (21)For the notion of isometry see also at p. 42.
Corrado Marastoni 33 Mathematical Analysis II - Part B
n The elements of n are called the Lebesgue-measurable subsets of R , and the restriction M n of ⇤ to n (coherently denoted by n)iscalledLebesgue measure on R . Lebesgue measure n M
n The family n of Lebesgue-measurable subsets of R is very large, so large that it is really M complicated to construct non measurable subsets: in practice, measurability is always verified.
Of particular importance is the role played by the subsets of Rn which are measurable with measure zero, called (Lebesgue-)null sets (or n-null sets). A property P(x) of the Null sets n points x of a subset E R is said to be true almost everywhere (often shortened into Almost ⇢ everywhere (a.e.) a.e.)inE when the set x E :P(x) is false is a null set. { 2 }
Example. A function f : E R is “continuous a.e. in E ”ifthesetofdiscontinuitypointsinE is a null ! set: so are e.g. the functions “entire part” and “fractional part” in R (the discontinuity points are those of Z,whichisanullsetinR). Proposition 2.1.3. The null sets have the following properties. (a) A countable union of null sets is a null set. (b) A subset of a null set is a null set. (c) The proper (i.e. of dimension n 1) submanifolds of Rn are null sets in Rn, as well as their finite or countable unions.
Examples. (1) AsinglepointisnullinR, and the same holds for the countable sets (as Z,orQ,orthe 2 Dirichlet set Q [0, 1]). (2) In R the single points and the regular curves are null, as well as their finite \ 2 or countable unions, are null. (3) In R the single points, the regular curves and the regular surfaces are null, as well as their finite or countable unions.
n A function f : R R is said to be measurable if the “over-level sets” Measurable function ! 1 n e f (]↵, + ]) = x R : f(x) >↵ 1 { 2 } are measurable for any ↵ R. 2 Since the family of measurable subset is very large, also the family of measurable functions is very large: in practice all functions are measurable. The following proposition shows that this property is stable enough.
Proposition 2.1.4. Measurable function enjoy the following properties. (a) The definition of measurability could be modified, without changing the class of functi- ons enjoying it, by substituting > with <, or , or also by requiring that the inverse image of any open subset (or of any closed subset) of R be measurable.
Corrado Marastoni 34 Mathematical Analysis II - Part B
(b) If f is measurable, also f , f + and f are so.(22) | | (c) If f and g are measurable, also f + g and fg are measurable.
(d) If fn is a sequence of measurable functions which converges pointwise to a function f, then also f is measurable.
The Lebesgue integral of a measurable function f : Rn R on a measurable set E Rn ! ⇢ is defined as follows. e If f : Rn R is measurable and 0 on E,thenthetrapezoid (see Figure 2.2(1)) • ! Trap (f)= (x, t) E R :0 t f(x) e E { 2 ⇥ } is a measurable subset of Rn+1, and the integral is defined in the following natural way (the value could be + ): 1
f(x) d n := n+1(TrapE (f)) . ZE
This is a simplified (but correct) version of the definition of Lebesgue integral. The usual definition uses the approximation of a positive measurable function by means of an increasing sequence of measurable simple functions, defined as those functions having only finitely many values; and the integral of a positive simple function is defined by summing on value strata. In fact Lebesgue theory, unlike Riemann’s, focuses the attention more on the codomain than on the domain (this was already visible in the previous definition of measurable function): loosely speaking, while Riemann divides the domain in intervals and then approximates the volume between the intervals and the graph by means of rectangles, Lebesgue divides the codomain in strata and then, for any level, measures the set of the elements of the domain whose image goes beyond that level. A simple pictorial representation of this idea can be seen in Figure 2.2(2). However, as one can imagine, in general such “over-level sets” could have a complicate structure, and to assign them a measure is much more challenging than assigning it to intervals: hence it is natural that the first interest of Lebesgue was to enlarge n as much as possible the class of measurable subsets of R .
n + If f : R R is any measurable function and at least one out of f (x) d n and • ! E f (x) d is finite, we set E n R e R + f(x) d := f (x) d f (x) d n n n ZE ZE ZE (this value could be + , or ). 1 1 + In particular, f is said to be (Lebesgue-)integrable on E when both E f (x) d n Integrable function 1 and f (x) d exist finite, and we shall write f L (E). Space L1(E) E n 2 R (22) + Recall thatR the positive part f and the negative part f of a function f are the functions defined as f +(x)=sup(f(x), 0) and f +(x)=sup( f(x), 0) (in other words, f + coincides with f where f is positive, and is zero elsewhere; and f coincides with f where f is negative, and is zero elsewhere). Note that both + + + f and f are positive functions (in spite of the name of f ), and that f = f f and f = f + f . | |
Corrado Marastoni 35 Mathematical Analysis II - Part B
1 Figure 2.2: (1) The trapezoid of f(x)=3+sinx over E =[ 2 , 3]. (2) The ideas of Riemann and Lebesgue integral.
Let us list some facts about Lebesgue integral. Beyond the expected properties (as the three —linearity, isotony, fundamental inequality— already known from Riemann’s), the most interesting features are the frequent presence of the condition “a.e.” (or “almost- everywhere”, which allows one to check that a hypothesis is satisfied up to null sets) and the theorems about the passage to the limit, which are more natural and e↵ective than the Riemann’s analogous ones.
Proposition 2.1.5. The following properties and results about Lebesgue integral hold true.
(a) (Countable additivity in the domain) Let Ek : k N be a countable family of { 2 } measurable pairwise disjoint subsets of Rn, f : Rn R a function, and denote ! E = E .Iff is measurable and positive, or if f L1(E), then k k 2 S fd n = fd n . E Ek Z Xk Z
(b) (Linearity) Given a measurable subset E of Rn, the set of functions L1(E) is a vector space on R, and the integral is a linear form on it. In other words: if f,g L1(E) and 2 ↵, R then ↵f + g L1(E),and 2 2
(↵f + g) d n = ↵ fd n + gd n . ZE ZE ZE (c) (Isotony) If f,g L1(E) and f g a.e. in E, then fd gd . 2 E n E n (d) (Fundamental inequality) Let f be measurable. Then f L1(E) if and only if R 2 R f L1(E) , and in such case it holds fd f d . | |2 E n E | | n 1 1 (e) (Comparison) Let f,g be measurable with R f g a.e.R . If g L (E) then f L (E). | | 2 2
Corrado Marastoni 36 Mathematical Analysis II - Part B
(f) (Integral and null sets) One has f =0 if and only if f is zero a.e. in E. E | | In general, two measurable functions which coincide a.e. have the same integral.(23) R (g) (Integral and boundedness) A function measurable and bounded a.e. (e.g.o, continuous on a compact subset) is integrable on any set of finite measure.
(h) (Monotone convergence for positive functions) Let E be measurable, and let fk be an increasing sequence of positive measurable functions. Then, called f the pointwise limit (24) of the fk, it holds fd n =limk + fk d n (could be + ). E ! 1 E 1 (i) (Monotone convergenceR for integrable functions)R Let E be measurable, and let fk be a monotone sequence of integrable functions. Then the sequence fk converges pointwise a.e. to a function f integrable on E if and only if the sequence E fd n is bounded; and in that case fd n =limk + fk d n R . E ! 1 E 2 R (l) (Dominated convergence)R Let E be measurable,R and be fk be a sequence of measurable functions which converges a.e. to a measurable function f. If there exists g L1(E) 1 2 such that fk g a.e., then f L (E) and fd n =limk + fk d n . | | 2 E ! 1 E (m) (Integration of series of positive functions) LetR E be measurable, andR let fk be a se- quence of positive measurable functions. Setting f(x)= k fk(x) (possibly having values in [0, + ]),itholds fd n = f d n . 1 E k E k P R P R We now describe more carefully the relation between Riemann and Lebesgue integrals.
Proposition 2.1.6. Riemann and Lebesgue integrals are related by the following facts. (a) (Integral on compact subsets) Let K be a elementarily measurable compact subset of Rn (e.g. an interval), and let f : K R.Iff is Riemann-integrable then f is also ! Lebesgue-integrable —i.e. f L1(K)— and the two integrals coincide. 2 Conversely, if f is measurable and bounded (hence in particular Lebesgue-integrable), then f is Riemann-integrable if and only if it is continuous a.e. in K. (b) (Generalized Riemann integral and Lebesgue integral) Let I be an interval of R and f : I R a locally Riemann-integrable function. Then f is Lebesgue-integrable on ! I if and only if it is absolutely Riemann-integrable in generalized sense on I,andin such case the Lebesgue integral of f on I is equal to the generalized Riemann integral of f on I. In particular, if f is Riemann-integrable in generalized sense on I but not absolutely,(25) then f is not Lebesgue-integrable on I. (c) (Fundamental Theorem of Calculus) If f :[a, b] R is Lebesgue-integrable, setting x ! F (x)= a fd 1 it holds F 0(x)=f(x) a.e. in [a, b].
(23)(Of course,R we mean, whenever the integral makes sense: hence when we are dealing with positive functions, or more generally when at least one of the integrals of positive and negative parts is finite.) The essential point of the statement is that, altering a measurable function only on a null subset of the domain (e.g. by setting there the function to zero), its integral does not change. In fact, for the Lebesgue integration two functions which coincide a.e. are indistinguishable: an evident example is the Dirichlet function , which is a.e. zero and hence has zero integral. Q [0,1] (24) \ The limit f(x):=limk fk(x)existsin[0, + ] for any x E,because(fk(x))k is positive and increasing. (25) sin x 1 2 for example, the function Dirichlet x on I = R.
Corrado Marastoni 37 Mathematical Analysis II - Part B
Conversely, let F :[a, b] R be derivable in any point of [a, b], and set f = F 0.Iff ! x is Lebesgue-integrable on [a, b], then it holds F (x) F (a)= fd . a 1 R
2.2 Integral calculus on a ne space
We now present the most relevant results for the applications in calculus, starting with a concrete example.
Let f(x, y)=x +2y and D = (x, y) R2 : x2 + y2 < 4 ,y>0 (see Figure 2.3). { 2 } Since f is continuous and D is measurable and bounded, the integral D f(x, y) d 2(x, y) (for which from now on we shall better use the more convenient notation f(x, y) dx dy) R D exists finite: let us compute it. R
Figure 2.3: (1) The set D,andthey-slice Dy at y =0,7. (2) The graph of f(x, y).
Now, Fubini’s theorem (see Proposition 2.2.1(a) below) says that this integral can be computed “by integrating on one variable at time”, or “by iterated integration”, by cutting the integration set D in slices where some variables are constant. For example, if we let y 2 2 vary from 0 to 2, then the “y-slice” of D is Dy = x R : 4 y x 4 y ,so that { 2 } p p 2 2 x=p4 y2 (x +2y) dx dy = dy (x +2y) dx = dy (x +2y) dx 2 D 0 Dy 0 x= p4 y Z Z Z Z Z 2 2 x=p4 y2 = ( 1 x2 +2xy] dy = ( 1 (4 y2)+2y 4 y2) ( 1 (4 y2) 2y 4 y2) dy 2 x= p4 y2 2 2 0 0 Z 2 Z 3 p p 2 4 2 2 2 32 32 = 4y 4 y dy =( 3 (4 y ) ]0 = (0) ( 3 )= 3 . 0 Z p
Corrado Marastoni 38 Mathematical Analysis II - Part B
Changing the order of integration would not have changed the result: if we let x vary from 2 2 to 2, then the “x-slice” of D is Dx = y R :0 y p4 x , so that { 2 } 2 2 y=p4 x2 (x +2y) dx dy = dx (x +2y) dy = dx (x +2y) dy D 2 D 2 y=0 Z Z Z x Z Z 2 2 2 y=p4 x2 2 2 = (xy + y ]y=0 dx = (x 4 x +4 x ) (0) dx 2 2 Z Z 1 2 3 1 3 2 p8 8 32 =( (4 x ) 2 +4x x ] =(8 ) ( 8+ )= . 3 3 0 3 3 3
But in this case the particular shape of D and of f suggests to use polar coordinates instead than the cartesian ones, i.e. (x, y)= (r, ✓)=(r cos ✓, r sin ✓). The theorem of change of variables (see Proposition 2.2.1(c) below) explains how to pass from a cartesian 1 integration to a polar one: it is enough to describe D in polar coordinates (i.e. (D)), compose f by , and also multiply by the absolute value of the determinant of the jacobian 0(r, ✓), which is r. Hence (using Fubini) we obtain once again
⇡ 2 (x +2y) dx dy = f( (r, ✓)) rdrd✓ = d✓ r2(cos ✓ +2sin✓) dr 1 ZD Z (D) Z0 Z0 ⇡ ⇡ 1 3 r=2 8 32 = ( 3 r (cos ✓ +2sin✓)]r=0 d✓ = 3 (cos ✓ +2sin✓) d✓ = 3 . Z0 Z0
Now we turn to precise statements, in the general case.
In the following we mean that n = n + n with n ,n N, and M N M N 2 n n x =(x ,x )withx =(x1,...,xn ) R M ,x=(xn +1,...,xn) R N ; M N M M 2 N M 2 n n n we shall denote the natural projections of R on R M and R N respectively with
n n n n ⇡ : R R M ,⇡(x)=x ,⇡: R R N ,⇡(x)=x ; M ! M M N ! N N n n moreover, given a subset E R and a function f : E R, for x R M we define the ⇢ ! M 2 x -section of E and of f setting respectively M
n Ex = x R N :(x ,x ) E ,fx : Ex R with fx (x )=f(x ,x ) M { N 2 M N 2 } M M ! M N M N
(analogous definitions are given for Ex and fx : Ex R). N N N ! Proposition 2.2.1. The following facts hold true. (a) (Fubini’s Reduction Theorem) Let E Rn be measurable and f an integrable function ⇢ on E. Then fx is integrable on Ex for a.e. x ⇡ (E); the function (defined a.e. M M M 2 M
Corrado Marastoni 39 Mathematical Analysis II - Part B
on ⇡ (E)) given by x f(x ,x ) d n (x ) is integrable on ⇡ (E);and M M Ex M N N N M 7! M R
fd n = f(x ,x ) d n (x ) d n (x ) . M N N N M M E ⇡ (E) Ex ! Z Z M Z M In particular we mean that all iterated integrals at the right-hand side, which could be obtained by varying the possible decompositions n = n + n and of the possible M N subfamilies of variables x and x give the same result. M N (b) (Tonelli’s Integrability Theorem) Let E Rn be measurable and f a measurable func- ⇢ tion on E. If any iterated integral of the modulus
f(x ,x ) d n (x ) d n (x ) M N N N M M ⇡ (E) Ex ! Z M Z M exists finite, then f is integrable on E (and for the computation one can apply Fubini). (c) (Change of variables) Let : B A be a di↵eomorphism between open subsets of ! Rn, and let E be a measurable subset of A. Given a function f : A R, we have that ! f L1(E) if and only if (f ) det L1( 1(E)) (where is the jacobian of 2 | 0|2 0 ),and
f(x) d n(x)= f( (⇠)) det 0(⇠) d n(⇠) . 1 | | ZE Z (E) Proof. Omitted.
Before providing some examples, let us make a couple of remarks.
Fubini’s Theorem is usually stated in the following alternative way (which we shall • n preferably use from now on), where the measure n on R is denoted by dx and where it is suggested that this measure could be interpreted as the “product measure” of n n the measures n = dx on M and n = dx on N : M M R N N R
f(x ,x ) dx dx = f(x ,x ) dx dx . M N M N M N N M E ⇡ (E) Ex ! Z Z M Z M Moreover, for the sake of simplicity it is customary to write the integral at the right-hand side also as
dx f(x ,x ) dx , M M N N ⇡ (E) Ex Z M Z M where we mean that f(x ,x ) is integrated first with respect to x on Ex (for a M N N M generic x ⇡ (E)) obtaining a function of x which will be then integrated on M 2 M M ⇡M(E).
Corrado Marastoni 40 Mathematical Analysis II - Part B
As we already said, on a set of finite measure any function measurable and bounded • a.e. (e.g. a continuous function) is Lebesgue-integrable. Beyond that case (hence when one deals with unbounded domains, or unbounded functions), the Theorem of Tonelli provides the most useful Lebesgue-integrability criterion: if an iterated integral of the modulus of the function exists finite then the function is integrable, and to compute its integral one can use Fubini’s Theorem. However one should be careful about the correct application of Tonelli’s result (hence by using the modulus), otherwise the conclusion could be false: some examples will be shown here below.
The most important changes of variables are the following ones. • 2 2 – Polar coordinates in R .If A = R (x, 0) : x 0 with coordinates (x, y), Polar \{ } coordinates and A =]0, + [ ] ⇡,⇡[ with coordinates (r, ✓), we set 0 1 ⇥ (x, y)= (r, ✓)=(r cos ✓, r sin ✓) .
cos ✓ r sin ✓ Therefore it holds 0(r, ✓)= sin ✓ rcos ✓ , and det 0 = r . ✓ ◆
Figure 2.4: (a) Polar coordinates in the plane. (b) Cylindrical and (c) spherical coordinates in the space.
3 3 – Cylindrical coordinates in R .If A = R (x, 0,z):x 0 with coordinates Cylindrical \{ } coordinates (x, y, z), and A0 =]0, + [ ] ⇡,⇡[ R with coordinates (r, ✓, z), we set 1 ⇥ ⇥ (x, y, z)= (r, ✓, z)=(r cos ✓, r sin ✓, z) .
cos ✓ r sin ✓ 0 Then it holds 0(r, ✓, z)= sin ✓rcos ✓ 0 , and det 0 = r . 0 0011 3@ 3 A – Spherical coordinates in R .If A = R (x, 0,z):x 0 with coordinates Spherical \{ } coordinates (x, y, z), and A =]0, + [ ] ⇡,⇡[ ]0,⇡[ with coordinates (r, ✓,'), we set 0 1 ⇥ ⇥ (x, y, z)= (r, ✓,')=(r cos ✓ sin ', r sin ✓ sin ', r cos ') .
cos ✓ sin ' r sin ✓ sin 'rcos ✓ cos ' 2 Then 0(r, ✓,')= sin ✓ sin 'rcos ✓ sin 'rsin ✓ cos ' , and det 0 = r sin ' . 0 cos ' 0 r sin ' 1 | | @ A
Corrado Marastoni 41 Mathematical Analysis II - Part B
(26) – A ne transformations, isometries, homotheties. In an euclidean space V , A ne an “a ne transformation” T : V V is a function of type T (v)=↵(v)+v , transformations ! 0 where ↵ : V V is a linear function (also called “linear part” of T )whilev is a ! 0 fixed vector which represents a translation. On the other hand an “isometry” is Isometries a self-di↵eomorphism : V V which preserves euclidean distances, i.e. such ! that (v) (w) = v w for any v, w V : it is known that isometries are || || || || 2 exactly the a ne transformations whose linear part is orthogonal (hence such t 1 (27) that ↵ = ↵ ) , which describes a rotation possibly followed by a reflection with respecto to an hyperplane of V . Another important example of a ne
transformation is given by the “homothety” of ratio µ>0, i.e. Tµ = µ idV :the e↵ect is clearly to dilate/contract all distances by a factor µ. In coordinates (thinking at V = Rn with the usual euclidean scalar product) an a ne transformation T : Rn Rn is therefore of type T (x)=Ax + b where b ! is a fixed vector of Rn and A is ae n n matrix: if E is a measurable bounded n ⇥ subset of R (hence its measure n(E) is finite), by the change of variables one has (T (E)) = det A (E) . The transformation is an isometry if and only n | | n if det A = 1, i.e. det A = 1: in this case the volume E is preserved. As for | | ⌥ n n the homothety of ratio µ one has det A = µ ,hence n(Tµ(E)) = µ n(E) (as it is reasonable to expect, if the distances are dilated by a factor µ then n-dimensional volumes are dilated by a factor µn).
Examples. (1) The function f(x, y)=3x2y 1 is continuous, hence integrable on any compact subset of 2 R : let us compute the integrals on the rectangle D1 =[ 1, 2] [0, 1] and on the triangle D2 of vertexes ⇥ (0, 0), (2, 0) and (1, 1). The projection of D1 on x is [ 1, 2], and for a generic x [ 1, 2] the x-section • 2 2 1 2 of D1 on y is always [0, 1]: therefore by Fubini one has f(x, y) dx dy = dx (3x y 1) dy = D1 1 0 2 3 2 2 y=1 2 3 2 1 3 x=2 1 3 1( 2 x y y]y=0 dx = 1( 2 x 1) dx =(2 x x]x=R 1 =(2) ( 2 )=R2 . InvertingR the order of integrations the result must be the same: namely the projection of D1 on y is [0, 1], and for a generic R R 1 2 2 y [0, 1] the y-section of D1 on x is always [ 1, 2], and we find f(x, y) dx dy = dy (3x y 2 D1 0 1 1 3 x=2 1 9 2 y=1 3 3 1) dx = (x y x]x= 1 dy = (9y 3) dy =( y 3y]y=0 =( ) (0) = . The projection of 0 0 2 R 2 2 • R R D2 on y is [0, 1], and for a generic y [0, 1] the y-section of D2 on x is [y, 2 y]: therefore by Fubini R 1 2 y R2 2 1 3 x=2 y 1 4 3 2 f(x, y) dx dy = dy (3x y 1) dx = (x y x]x=y dy = ( 2y +6y 12y +10y 2) dy = D2 0 y 0 0 2 5 3 4 3 2 1 1 ( y + y 4y +5y 2y] = . Also in this case let us invert the order. The projection of D2 on x is R 5 2 R R 0 10 R R [0, 2]; for a generic x [0, 1] the x-section of D2 on y is [0,x], while for a generic x [1, 2] is [0, 2 x]: we 2 1 x 2 2 2 x 2 21 3 2 2 y =x then obtain f(x, y) dx dy = dx (3x y 1) dy + dx (3x y 1) dy = ( x y y] dx + D2 0 0 1 0 0 2 y=0 2 3 2 2 y=2 x 1 3 4 2 3 2 2 ( x y y] dx = ( x x) dx + ( x (2 x) (2 x)) dx, which eventually gives once more 1 2 R y=0 0 2 R R 1 2 R R R 1 + 3 = 1 . (2) (Area of a plane region described in polar coordinates) Let us consider a region of R 5 10 10 R R the cartesian plane in polar coordinates given by D = (⇢,✓):✓ [↵, ],⇢1(✓) ⇢ ⇢2(✓) for certain { 2 } 0 ↵<