9.4: Convergence of Parametrized Integrals and Functions
- Last updated
- Jan 15, 2020
- Save as PDF
- Page ID
- 32391
( \newcommand{\kernel}{\mathrm{null}\,}\)
I. We now consider C-integrals of the form
C∫f(t,u)dm(t),
where m is Lebesgue or LS measure in E1. Here the variable u, called a parameter, remains fixed in the process of integration; but the end result depends on u, of course.
We assume f:E2→E (E complete) even if not stated explicitly. As before, we give our definitions and theorems for the case
C∫∞a.
The other cases (C∫a−∞,C∫b−a, etc. ) are analogous; they are treated in Problems 2 and 3. We assume
a,b,c,x,t,u,v∈E1
throughout, and write "dt" for "dm(t)" iff m is Lebesgue measure.
If
C∫∞af(t,u)dm(t)
converges for each u in a set B⊆E1, we can define a map F:B→E by
F(u)=C∫∞af(t,u)dm(t)=limx→∞∫xaf(t,u)dm(t).
This means that
(∀u∈B)(∀ε>0)(∃b>a)(∀x≥b)|∫xaf(t,u)dm(t)−F(u)|<ε,
so |F|<∞ on B.
Here b depends on both ε and u (convergence is "pointwise"). However, it may occur that one and the same b fits all u∈B, so that b depends on ε alone. We then say that
C∫∞af(t,u)dm(t)
converges uniformly on B (i.e., for u∈B), and write
F(u)=C∫∞af(t,u)dm(t) (uniformly) on B.
Explicitly, this means that
(∀ε>0)(∃b>a)(∀u∈B)(∀x≥b)|∫xaf(t,u)dm(t)−F(u)|<ε.
Clearly, this implies (1), but not conversely. We now obtain the following.
Theorem 9.4.1 (Cauchy criterion)
Suppose
∫xaf(t,u)dm(t)
exists for x≥a and u∈B⊆E1. (This is automatic if E⊆E∗; see Chapter 8, §5.)
Then
C∫∞af(t,u)dm(t)
converges uniformly on B iff for every ε>0, there is b>a such that
(∀v,x∈[b,∞))(∀u∈B)|∫xvf(t,u)dm(t)|<ε,
and
|∫baf(t,u)dm(t)|<∞.
- Proof
-
The necessity of (3) follows as in Theorem 2 of §3. (Verify!)
To prove sufficiency, suppose the desired b exists for every ε>0. Then for each (fixed) u∈B,
C∫∞af(t,u)dm(t)
satisfies Theorem 2 of §3. Hence
F(u)=limx→∞∫xaf(t,u)dm(t)≠±∞
exists for every u∈B (pointwise). Now, from (3), writing briefly ∫f for ∫f(t,u)dm(t), we obtain
|∫xvf|=|∫xaf−∫vaf|<ε
for all u∈B and all x>v≥b.
Making x→∞ (with u and v temporarily fixed), we have by (4) that
|F(u)−∫vaf|≤ε
whenever v≥b.
But by our assumption, b depends on ε alone (not on u). Thus unfixing u, we see that (5) establishes the uniform convergence of
∫∞af,
as required.◻
Corollary 9.4.1
Under the assumptions of Theorem 1,
C∫∞af(t,u)dm(t)
converges uniformly on B if
C∫∞a|f(t,u)|dm(t)
does.
Indeed,
|∫xvf|≤∫xv|f|<ε.
Corollary 9.4.2 (comparison test)
Let f:E2→E and M:E2→E∗ satisfy
|f(t,u)|≤M(t,u)
for u∈B⊆E1 and t≥a.
Then
C∫∞a|f(t,u)|dm(t)
converges uniformly on B if
C∫∞aM(t,u)dm(t)
does.
Indeed, Theorem 1 applies, with
|∫xvf|≤∫xvM<ε.
Hence we have the following corollary.
Corollary 9.4.3 ("M-test")
Let f:E2→E and M:E1→E∗ satisfy
|f(t,u)|≤M(t)
for u∈B⊆E1 and t≥a. Suppose
C∫∞aM(t)dm(t)
converges. Then
C∫∞a|f(t,u)|dm(t)
converges (uniformly) on B. So does
C∫∞af(t,u)dm(t)
by Corollary 1.
- Proof
-
Set
h(t,u)=M(t)≥|f(t,u)|.
Then Corollary 2 applies (with M replaced by h there). Indeed, the convergence of
C∫h=C∫M
is trivially "uniform" for u∈B, since M does not depend on u at all.◻
Note 1. Observe also that, if h(t,u) does not depend on u, then the (pointwise) and (uniform) convergence of C∫h are trivially equivalent.
We also have the following result.
Corollary 9.4.4
Suppose
C∫∞af(t,u)dm(t)
converges (pointwise) on B⊆E1. Then this convergence is uniform iff
limν→∞C∫∞vf(t,u)dm(t)=0 (uniformly) on B,
i.e., iff
(∀ε>0)(∃b>a)(∀u∈B)(∀v≥b)|C∫∞vf(t,u)dm(t)|<ε.
- Proof
-
The proof (based on Theorem 1) is left to the reader, along with that of the following corollary.
Corollary 9.4.5
Suppose
∫baf(t,u)dm(t)≠±∞
exists for each u∈B⊆E1.
Then
C∫∞af(t,u)dm(t)
converges (uniformly) on B iff
C∫∞bf(t,u)dm(t)
does.
II. The Abel-Dirichlet tests for uniform convergence of series (Problems 9 and 11 in Chapter 4, §13) have various analogues for C-integrals. We give two of them, using the second law of the mean (Corollary 5 in §1).
First, however, we generalize our definitions, "unstarring" some ideas of Chapter 4, §11. Specifically, given
H:E2→E (E complete),
we say that H(x,y) converges to F(y), uniformly on B, as x→q(q∈E∗), and write
limx→qH(x,y)=F(y) (uniformly) on B
iff we have
(∀ε>0)(∃G¬q)(∀y∈B)(∀x∈G¬q)|H(x,y)−F(y)|<ε;
hence |F|<∞ on B.
If here q=∞, the deleted globe G¬q has the form (b,∞). Thus if
H(x,u)=∫xaf(t,u)dt,
(6) turns into (2) as a special case. If (6) holds with "(∃G¬q)" and "(∀y∈B)" interchanged, as in (1), convergence is pointwise only.
As in Chapter 8, §8, we denote by f(⋅,y), or fy, the function of x alone (on E1) given by
fy(x)=f(x,y).
Similarly,
fx(y)=f(x,y).
Of course, we may replace f(x,y) by f(t,u) or H(t,u), etc.
We use Lebesgue measure in Theorems 2 and 3 below.
Theorem 9.4.2
Assume f,g:E2→E1 satisfy
(i) C∫∞ag(t,u)dt converges (uniformly) on B;
(ii) each gu(u∈B) is L-measurable on A=[a,∞);
(iii) each fu(u∈B) is monotone (↓ or ↑) on A; and
(iv) |f|<K∈E1 (bounded) on A×B.
Then
C∫∞af(t,u)g(t,u)dt
converges uniformly on B.
- Proof
-
Given ε>0, use assumption (i) and Theorem 1 to choose b>a so that
|L∫xvg(t,u)dt|<ε2K,
written briefly as
|L∫xvgu|<ε2K,
for all u∈B and x>v≥b, with K as in (iv).
Hence by (ii), each gu(u∈B) is L-integrable on any interval [v,x]⊂A, with x>v≥b. Thus given such u and [v,x], we can use (iii) and Corollary 5 from §1 to find that
L∫xvfugu=fu(v)L∫cvgu+fu(x)L∫xcgu
for some c∈[v,x].
Combining with (7) and using (iv), we easily obtain
|L∫xvf(t,u)g(t,u)dt|<ε
whenever u∈B and x>v≥b. (Verify!)
Our assertion now follows by Theorem 1.◻
Theorem 9.4.3 (Abel-Dirichlet test)
Let f,g:E2→E∗ satisfy
(a) limt→∞f(t,u)=0 (uniformly) for u∈B;
(b) each fu(u∈B) is nonincreasing (↓) on A=[0,∞);
(c) each gu(u∈B) is L-measurable on A; and
(d) (∃K∈E1)(∀x∈A)(∀u∈B)|L∫xag(t,u)dt|<K.
Then
C∫∞af(t,u)g(t,u)dt
converges uniformly on B.
- Proof Outline
-
Argue as in Problem 13 of §3, replacing Theorem 2 in §3 by Theorem 1 of the present section.
By Lemma 2 in §1, obtain
|L∫xvfugu|=|fu(v)L∫xagu|≤Kf(v,u)
for u∈B and x>v≥a.
Then use assumption (a) to fix k so that
|f(t,u)|<ε2K
for t>k and u∈B.◻
Note 2. Via components, Theorems 2 and 3 extend to the case g:E2→ En(Cn).
Note 3. While Corollaries 2 and 3 apply to absolute convergence only, Theorems 2 and 3 cover conditional convergence, too (a great advantage!). The theorems also apply if f or g is independent of u (see Note 1). This supersedes Problems 13 and 14 in §3.
Examples
(A) The integral
∫∞0sintutdt
converges uniformly on Bδ=[δ,∞) if δ>0, and pointwise on B=[0,∞).
Indeed, we can use Theorem 3, with
g(t,u)=sintu
and
f(t,u)=1t,f(0,u)=1,
say. Then the limit
limt→∞1t=0
is trivially uniform for u∈Bδ, as f is independent of u. Thus assumption (a) is satisfied. So is (d) because
|∫x0sintudt|=|1u∫xu0sinθdθ|≤1δ⋅2.
(Explain!) The rest is easy.
Note that Theorem 2 fails here since assumption (i) is not satisfied.
(B) The integral
∫∞01te−tusinatdt
converges uniformly on B=[0,∞). It does so absolutely on Bδ=[δ,∞), if δ>0.
Here we shall use Theorem 2 (though Theorem 3 works, too). Set
f(t,u)=e−tu
and
g(t,u)=sinatt,g(0,u)=a.
Then
∫∞0g(t,u)dt
converges (substitute x=at in Problem 8 or 15 in §3). Convergence is trivially uniform, by Note 1. Thus assumption (i) holds, and so do the other assumptions. Hence the result.
For absolute convergence on Bδ, use Corollary 3 with
M(t)=e−δt,
so M≥|fg|.
Note that, quite similarly, one treats C-integrals of the form
∫∞ae−tug(t)dt,∫∞ae−t2ug(t)dt, etc.,
provided
∫∞ag(t)dt
converges (a≥0).
In fact, Theorem 2 states (roughly) that the uniform convergence of C∫g implies that of C∫fg, provided f is monotone (in t) and bounded.
III. We conclude with some theorems on uniform convergence of functions H:E2→E (see (6)). In Theorem 4, m is again an LS (or Lebesgue) measure in E1; the deleted globe G∗¬q is fixed.
Theorem 9.4.4
Suppose
limx→qH(x,y)=F(y) (uniformly)
for y∈B⊆E1. Then we have the following:
(i) If all Hx(x∈G∗¬q) are continuous or m-measurable on B, so also is F.
(ii) The same applies to m-integrability on B, provided mB<∞; and then
limx→q∫B|Hx−F|=0;
hence
limx→q∫BHx=∫BF=∫B(limx→qHx).
Formula (8') is known as the rule of passage to the limit under the integral sign.
- Proof
-
(i) Fix a sequence xk→q (xk in the deleted globe G∗¬q), and set
Hk=Hxk(k=1,2,…).
The uniform convergence
H(x,y)→F(y)
is preserved as x runs over that sequence (see Problem 4). Hence if all Hk are continuous or measurable, so is F (Theorem 2 in Chapter 4, §12 and Theorem 4 in Chapter 8, §1. Thus clause (i) is proved.
(ii) Now let all Hx be m-integrable on B; let
mB<∞.
Then the Hk are m-measurable on B, and so is F, by (i). Also, by (6),
(∀ε>0)(∃G¬q)(∀x∈G¬q)∫B|Hx−F|≤∫B(ε)=εmB<∞,
proving (8). Moreover, as
∫B|Hx−F|<∞,
Hx−F is m-integrable on B, and so is
F=Hx−(Hx−F).
Hence
|∫BHx−∫BF|=|∫B(Hx−F)|≤∫B|Hx−F|→0,
as x→q, by (8). Thus (8') is proved, too.◻
Quite similarly (keeping E complete and using sequences), we obtain the following result.
Theorem 9.4.5
Suppose that
(i) all Hx(x∈G∗−q) are continuous and finite on a finite interval B⊂E1, and differentiable on B−Q, for a fixed countable set Q;
(ii) limx→qH(x,y0)≠±∞ exists for some y0∈B; and
(iii) limx→qD2H(x,y)=f(y) (uniformly) exists on B−Q.
Then f, so defined, has a primitive F on B, exact on B−Q (so F′=f on B−Q); moreover,
F(y)=limx→yH(x,y) (uniformly) for y∈B.
- Outline of Proof
-
Note that
D2H(x,y)=ddyHx(y).
Use Theorem 1 of Chapter 5, §9, with Fn=Hxn,xn→q.◻
Note 4. If x→q over a path P (clustering at q), one must replace G¬q and G∗¬q by P∩G¬q and P∩G∗¬q in (6) and in Theorems 4 and 5.