The theory of absolute and conditional convergence for integrals over an infinite range, and integration by parts as the integral version of Dirichlet’s test, with ∫sinx/xsdx as the main example.
Infinite integrals (§§184–189) were studied in the first part of this chapter for positive integrands, where the only question is whether the area is finite. But integrals such as ∫0∞(sinx)/xdx, which turns out to equal π/2 and is one of the most famous integrals in analysis, have integrands that change sign for ever. §203 transfers the whole theory of the last few lessons to integrals: absolute and conditional convergence, and a test in the spirit of Dirichlet’s, with integration by parts playing the role of summation by parts.
The idea
Picture ∫1X(sinx)/xdx as signed area: the arches of the curve above the axis count positively, those below negatively. Each arch spans an interval of length π, and their heights shrink like 1/x. So the arches have areas roughly 2/π,−2/(2π),2/(3π),…: in effect an alternating series with terms decreasing to 0, which converges. But the total unsigned area is like ∑2/(kπ), a harmonic series, which diverges.
ExploreArches of shrinking height
Try this.Keep s=1 and push X out to 200. Watch the two curves in the lower plot. Then raise s above 1, and lower it towards 0.1. Finally switch the lower limit to 0 and take s close to 2.
∫1Xxsinxdx
xsin x / xˢ
Xrunning integrals
top: positive area · bottom: F(X) = ∫ₐˣ f
top: negative area · bottom: ∫ₐˣ |f|, total unsigned area
value of the integral
Lower limit
Up to X=60: ∫1Xf=0.64066 and ∫1X∣f∣=2.7709. The lobes shrink and alternate, so F(X) settles, but their sizes are like 2/(kπ)s, whose sum diverges: ∫∣f∣ grows without bound. The integral converges only conditionally. Its value is π/2−Si(1)≈0.62471.
What to notice
For s=1 the blue curve F(X) wobbles around a limit and settles, while the ochre curve, the total unsigned area, keeps climbing like logX: conditional convergence. For s>1 both level off: absolute convergence. Smaller s makes the arches shrink more slowly, but F(X) still settles for every s>0. From 0, the integrand behaves like x1−s near 0, so the area there becomes infinite as s reaches 2.
Precise statements
DefinitionAbsolute and conditional convergence of integralsHardy: §203
The integral ∫a∞f(x)dx is absolutely convergent if ∫a∞∣f(x)∣dx converges. If ∫a∞f(x)dx converges but ∫a∞∣f(x)∣dx does not, it is conditionally convergent.
As for series, split f into its positive and negative parts, g(x)=21(∣f(x)∣+f(x)) and h(x)=21(∣f(x)∣−f(x)). Then g,h≥0, f=g−h, ∣f∣=g+h, and g, h are continuous where f is.
If ∫a∞∣f∣ converges, then ∫a∞g and ∫a∞h both converge, and so does ∫a∞f=∫a∞g−∫a∞h. If ∫a∞f converges only conditionally, then ∫a∞g and ∫a∞h both diverge. In particular, if ∣f(x)∣≤ϕ(x) and ∫a∞ϕ converges, then ∫a∞f is absolutely convergent.
The proof is that of §191 and §193 with integrals in place of sums (0≤g,h≤∣f∣, and comparison for positive integrands, §185). The interesting theorem is the analogue of Dirichlet’s test.
TheoremIntegrals with an oscillating factorHardy: §203
Suppose that ϕ′(x) is continuous for x≥a, that ϕ(x)≥0 and ϕ′(x)≤0, and that ϕ(x)→0 as x→∞. Then for every real t=0 the integrals
∫a∞ϕ(x)costxdx,∫a∞ϕ(x)sintxdx
are convergent.
CorollaryThe integrals of costx/xs and sintx/xsHardy: §203
For a>0 and t=0, ∫a∞x−scostxdx and ∫a∞x−ssintxdx are absolutely convergent if s>1 and conditionally convergent if 0<s≤1.
(Hardy’s text says “absolutely convergent when s>0”; the context and the companion clause show that s>1 is meant.)
Proofs
ProofIntegrals with an oscillating factor
Strategy
Integrate by parts to move the oscillating factor onto ϕ: the derivative ϕ′ is of one sign and has a finite integral, so the new integral converges absolutely, and the boundary term tends to 0. This is summation by parts in continuous form.
∫a∞∣ϕ′∣ converges
Since ϕ′≤0, ∣ϕ′∣=−ϕ′, and ∫aX∣ϕ′(x)∣dx=ϕ(a)−ϕ(X)→ϕ(a). (This plays the part of ∑(φk−φk+1)=φ0 in Dirichlet’s test.)
Integrate by parts
We may take t>0 (replace t by −t otherwise). On [a,X], with dxdtsintx=costx,
sin(tX)ϕ(X)/t≤ϕ(X)/t→0 as X→∞. The middle term does not depend on X.
The last integral converges
∣ϕ′(x)sintx∣≤∣ϕ′(x)∣, so by the first step and comparison, ∫a∞ϕ′(x)sintxdx is absolutely convergent, hence convergent.
Conclude
Every term on the right has a limit as X→∞, so the left-hand side does: ∫a∞ϕ(x)costxdx converges. The sine integral is the same with dxd(−tcostx)=sintx.
∎
ProofThe corollary: when is the convergence absolute?
Strategy
For s>1 dominate by x−s. For 0<s≤1, convergence comes from the theorem, and non-absolute convergence from ∣sintx∣≥sin2tx=21(1−cos2tx), exactly as in Examples LXXIX, 6 for series.
s>1
∣x−ssintx∣≤x−s, and ∫a∞x−sdx converges for s>1.
0<s≤1: convergence
ϕ(x)=x−s satisfies the hypotheses: ϕ′=−sx−s−1 is continuous and negative, and ϕ→0.
0<s≤1: not absolutely
∫aXxs∣sintx∣dx≥∫aXxssin2txdx=21∫aXxsdx−21∫aXxscos2txdx. The last integral converges by the theorem, while ∫aXx−sdx→∞ for s≤1. The same works for ∣costx∣≥cos2tx=21(1+cos2tx).
∎
Worked examples and exercises
ExerciseFrom zero to infinity●●○Hardy: Examples LXXXII, 1
Let t>0. Show that ∫0∞xssintxdx is convergent if 0<s<2, and absolutely convergent if 1<s<2.
Hint
Consider (0,1) and (1,∞) separately, as Hardy suggests.
Hint
Near 0, sintx is close to tx.
Solution
Near 0. The integrand is positive on (0,π/t) and sintx/xs∼tx1−s as x→0. Since 21tx≤sintx≤tx for small x>0, and ∫01x1−sdx converges exactly when 1−s>−1, that is s<2, the integral near 0 converges (absolutely) for s<2 and diverges for s≥2. On the rest of (0,1] the integrand is continuous.
On [1,∞). By the corollary: convergent for s>0, absolutely for s>1, and for s≤0 not convergent (the integral over each interval [2kπ/t,(2k+1)π/t] is at least 2/t in size when s≤0, so Cauchy’s principle fails).
Together: convergent for 0<s<2, absolutely for 1<s<2. In the figure, switch the lower limit to 0.
Let t=0. Show that ∫0∞xs1−costxdx converges if and only if 1<s<3, and that it is then absolutely convergent.
Hint
The integrand is never negative. What are its sizes near 0 and near ∞?
Solution
Since 1−costx≥0, convergence and absolute convergence are the same thing. Near 0: 1−costx∼21t2x2, so the integrand behaves like 21t2x2−s, integrable on (0,1) iff 2−s>−1, that is s<3. Near ∞: ∫1∞xs1−costxdx=∫1∞x−sdx−∫1∞xscostxdx; for s>1 both converge, and for 0<s≤1 the first diverges while the second converges, so the whole diverges; for s≤0 it diverges too, since the integrand is at least 1−costx, whose integral over each period is 2π/∣t∣. Hence the condition is 1<s<3.
Show that ∫0∞xssinx(1−cosx)dx is convergent if 0<s<4, and absolutely convergent if 1<s<4.
Hint
Near 0 the numerator is about 21x3. Near ∞ write sinx(1−cosx)=sinx−21sin2x.
Solution
Near 0.sinx(1−cosx)∼x⋅21x2, and it is positive on (0,π), so ∫01 converges iff 3−s>−1, that is s<4.
Near ∞.∫1∞xssinx−21sin2xdx is the difference of two integrals that converge for s>0 (the theorem), and ∣sinx(1−cosx)∣≤2 gives absolute convergence for s>1. For 0<s≤1 it is not absolute: p(x)=∣sinx(1−cosx)∣ is continuous, 2π-periodic and not identically 0, so c=∫02πp>0, and ∫2kπ2(k+1)πp(x)x−sdx≥c(2(k+1)π)−s, whose sum over k diverges. For s≤0 the integral over [2kπ,2kπ+π] of the integrand is at least ∫0πsinx(1−cosx)dx=2 in size, so it does not converge.
ExerciseAn integrand that oscillates near a point●●○Hardy: §190, §203
Show that ∫aAsin(x−a1)x−adx (with A>a) converges, but not absolutely.
Hint
Substitute y=1/(x−a).
Solution
With y=1/(x−a), dx/(x−a)=−dy/y, and x→a+ corresponds to y→∞. So, for a<ξ<A,
∫ξAsin(x−a1)x−adx=∫1/(A−a)1/(ξ−a)ysinydy.
As ξ→a the right-hand side tends to ∫1/(A−a)∞(siny)/ydy, which converges conditionally by the corollary (s=1). The same substitution with ∣sin∣ shows that the integral of the modulus diverges. An integral of the second kind, oscillating near a point, is thus turned into one of the first kind, oscillating at infinity (§186).
ExerciseChanging the variable●●●Hardy: Examples LXXXII, 5
Show that ∫0∞x−αsinx1−βdx converges if and only if α lies strictly between β and 2−β.
Hint
Put y=x1−β, treating β<1 and β>1 separately. You should arrive at ∫0∞y−σsinydy times a constant.
Solution
β<1. Then y=x1−β increases from 0 to ∞, x=y1/(1−β), dx=1−β1yβ/(1−β)dy, and
By Examples LXXXII, 1 this converges iff 0<σ<2, i.e. 0<α−β<2−2β, i.e. β<α<2−β.
β>1. Now y=x1−β decreases from ∞ to 0; the same computation gives a constant times ∫0∞y−σsinydy with σ=(β−α)/(β−1), convergent iff 0<β−α<2β−2, i.e. 2−β<α<β.
β=1. The integrand is x−αsin1, and ∫0∞x−αdx never converges; the interval between β and 2−β is then empty. In all cases: convergence iff α lies strictly between β and 2−β. (The substitution rule for infinite integrals is §186.)
Check yourself
Check yourselfTrue or false?
∫1∞(sinx)/xdx is absolutely convergent.
Why
False. It converges, but ∫1X∣sinx∣/xdx≥∫1Xsin2x/xdx, which grows like 21logX. Convergence is conditional.
Check yourselfChoose one
For which s does ∫0∞(sinx)/xsdx converge?
Why
At infinity it needs s>0 (Dirichlet-type theorem); near 0 the integrand is about x1−s, which needs s<2. Absolute convergence needs, in addition, s>1.
Check yourselfTrue or false?
In the proof that ∫a∞ϕ(x)costxdx converges, the only place where ϕ′≤0 is used is to show that ∫a∞∣ϕ′∣ is finite.
Why
True. The boundary term needs only ϕ→0; the remaining integral needs ∫∣ϕ′∣<∞, which for a decreasing ϕ is just ϕ(a)−limϕ. (The theorem stays true for any ϕ→0 of “bounded variation”, ∫∣ϕ′∣<∞.)
Hardy’s language
There is a theory for integrals analogous to that developed for series.
— G. H. Hardy, §203
Summary
Summary
∫a∞f is absolutely convergent if ∫a∞∣f∣ converges; then it converges, and its positive and negative parts converge separately.
If ϕ≥0 decreases to 0 (with ϕ′ continuous), then ∫a∞ϕ(x)costxdx and ∫a∞ϕ(x)sintxdx converge: integrate by parts, as summation by parts proves Dirichlet’s test.
∫a∞x−ssintxdx (a>0) converges absolutely for s>1 and conditionally for 0<s≤1; from 0, convergence also needs s<2.
Non-absolute convergence is shown by ∣sin∣≥sin2=21(1−cos2x), which splits off a divergent positive part.