Summation by parts turns bounded partial sums and steadily decreasing factors into convergence; the alternating series test, the series ∑φncosnθ, and Abel’s test all follow.
The alternating series test handles the sign pattern +,−,+,−,… only. But the series that matter most in later analysis, such as ∑(cosnθ)/n and ∑(sinnθ)/n, the building blocks of Fourier series, have signs that follow no simple pattern. §196 gives a test that needs much less: the factor carrying the signs must have bounded partial sums, and the factor carrying the sizes must decrease steadily to zero. The tool behind it, summation by parts, is the discrete version of integration by parts and will reappear in §203 for integrals.
The idea
Think of ∑anφn as a product of two factors with different jobs. The an (for instance cosnθ) supply the cancellation: their running totals An=a1+⋯+an never wander far from 0, although the series ∑an itself need not converge. The φn (for instance 1/n) supply the damping: they shrink steadily to 0.
Summation by parts rewrites the sum so that it uses only the totals An and the dropsφn−φn+1:
n=1∑Nanφn=n=1∑N−1An(φn−φn+1)+ANφN.
The drops are positive and add up to φ1−φN<φ1; the An are bounded. So the right-hand side is a series that converges absolutely, plus a last term that tends to 0. The cancellation hidden in the an has been converted into absolute convergence.
ExploreBounded totals times shrinking factors
Try this.With an=cosnθ and φn=1/n, move θ and watch both plots. Then take θ close to 0 (or close to 2π). Then change an to 1,1,1,…, and try φn=1/log(n+1) with the other factors.
∑anφn with an=cosnθ,φn=n1Partial sums An=a1+⋯+an of the factor an
nAₙ
Partial sums of ∑anφn
nΣ aₖφₖ
Aₙ
band ±K
partial sums of Σ aₙφₙ
room left for the tail: ±2Kφₙ₊₁
the sum
Every ∣An∣≤K=2.086 (K=1/∣sin21θ∣). Summation by parts then gives, for every m, ∣amφm+⋯+anφn∣≤2Kφm. After 200 terms the partial sum is 0.039228, and the rest of the series can move it by at most 2Kφ201=0.0208. The sum is 0.0420195.
What to notice
As long as the ochre totals An stay inside the green band ±K, the blue partial sums settle, and the purple bar, of half-length 2Kφn+1, always contains everything that is still to come. Near θ=0 the band widens like 2/θ and convergence is slow. With an=1 there is no band, and the blue curve climbs for ever.
Precise statements
Following Hardy we index from 0 here, write sn=a0+a1+⋯+an, and say that φntends steadily to zero if it is positive, decreasing and has limit 0.
LemmaAbel’s summation by partsHardy: §196
For any numbers a0,…,an and φ0,…,φn, with sk=a0+⋯+ak,
If φn tends steadily to zero, and ∑an converges or oscillates finitely (that is, its partial sums are bounded), then ∑anφn converges. Moreover, if ∣sn∣≤K for all n, then for every m≤n
∣amφm+⋯+anφn∣≤2Kφm.
With an=(−1)n, whose partial sums are 1,0,1,0,…, this is exactly the alternating series test of §195.
TheoremAbel’s testHardy: §196
If φn is positive and decreasing (its limit need not be 0), and ∑an converges, then ∑anφn converges. In Hardy’s words: a convergent series stays convergent when its terms are multiplied by any sequence of positive decreasing factors.
If θ is not a multiple of 2π, the partial sums of ∑n≥1cosnθ and ∑n≥1sinnθ are bounded by 1/∣sin21θ∣. Hence, if φn tends steadily to zero, ∑φncosnθ and ∑φnsinnθ converge for every θ, except perhaps the cosine series when θ is a multiple of 2π (where it becomes ∑φn).
Proofs
The identity is pure algebra: substitute ak=sk−sk−1 (with s−1=0) on the left and collect the coefficient of each sk; it is φk−φk+1 for k<n and φn for k=n. Exercise 1 asks you to check it.
ProofDirichlet’s test
Strategy
Use summation by parts. The new series ∑sk(φk−φk+1) has bounded sk times positive drops whose total is finite, so it converges absolutely; the leftover term snφn tends to 0.
Bound the partial sums
Because ∑an converges or oscillates finitely, its partial sums are bounded: there is K with ∣sk∣≤K for all k.
The drops form a convergent positive series
φk−φk+1≥0, and (φ0−φ1)+⋯+(φn−1−φn)=φ0−φn→φ0. So ∑(φk−φk+1) converges.
∑sk(φk−φk+1) converges absolutely
∣sk(φk−φk+1)∣≤K(φk−φk+1): comparison with the previous step. So, by §191, its partial sums tend to a limit.
The last term vanishes
∣snφn∣≤Kφn→0.
Conclude
By the identity, the partial sums of ∑anφn are the partial sums of a convergent series plus a term tending to 0; so they converge.
The tail bound
Apply the identity to the block am,…,an with the partial sums sk−sm−1, which are bounded by 2K. The same estimate gives at most 2K(φm−φn)+2Kφn=2Kφm.
∎
ProofAbel’s test from Dirichlet’s
Strategy
A decreasing positive φn has a limit l≥0. Then φn−l tends steadily to 0, and Dirichlet’s test applies to ∑an(φn−l).
φn→l
φn is decreasing and bounded below by 0, so it has a limit l≥0 (§69), and φn−l decreases steadily to 0.
Apply Dirichlet
∑an converges, so its partial sums are bounded, and ∑an(φn−l) converges.
Add back
∑anφn=∑an(φn−l)+l∑an, a sum of two convergent series.
∎
ProofThe bound for ∑cosnθ and ∑sinnθ
Strategy
Add the two series as one complex geometric series ∑zn with z=cosθ+isinθ (§88).
Sum the geometric series
With z=1, z+z2+⋯+zn=1−zz(1−zn), whose modulus is at most ∣1−z∣2 since ∣z∣=1.
Take real and imaginary parts
By De Moivre, the real and imaginary parts of the left-hand side are ∑k≤ncoskθ and ∑k≤nsinkθ; each is at most the modulus of the whole.
Compute ∣1−z∣
∣1−z∣2=(1−cosθ)2+sin2θ=2−2cosθ=4sin221θ, so 2/∣1−z∣=1/∣sin21θ∣.
∎
Worked examples and exercises
ExerciseChecking summation by parts●○○Hardy: §196
Verify Abel’s identity for n=2, and explain why it is the discrete analogue of ∫abfg=[Fg]ab−∫abFg′, where F′=f.
Hint
Expand the right-hand side using s0=a0, s1=a0+a1, s2=a0+a1+a2.
The partial sums sk play the part of the antiderivative F of f=a, the drops φk−φk+1 play the part of −g′, and snφn is the boundary term [Fg] (the lower boundary term vanishes because s−1=0). Dirichlet’s test is the discrete form of the argument used in §203 for ∫φ(x)costxdx.
ExerciseThe bare trigonometric series●○○Hardy: Examples LXXIX, 2
Show that ∑cosnθ and ∑sinnθ oscillate finitely when θ is not a multiple of π. What happens when θ is a multiple of π?
Hint
Bounded partial sums are the corollary. For non-convergence, look at the terms.
Solution
By the corollary the partial sums are bounded. The series do not converge because their terms do not tend to 0: if cosnθ→0 then cos2nθ=2cos2nθ−1→−1, a contradiction; and if sinnθ→0 then cos2nθ→1, so ∣sin(n+1)θ∣=∣sinnθcosθ+cosnθsinθ∣→∣sinθ∣=0, again a contradiction. So both oscillate finitely.
If θ=kπ: the sine series is 0+0+⋯ and converges to 0. The cosine series is ∑(−1)kn: for odd k it is −1+1−1+⋯ (finite oscillation), and for even k it is 1+1+⋯ (divergence to ∞).
ExerciseThe series ∑n−scosnθ●●○Hardy: Examples LXXIX, 3–4
Classify ∑n≥1n−scosnθ and ∑n≥1n−ssinnθ: show that they are absolutely convergent if s>1, convergent if 0<s≤1 (except the cosine series when θ is a multiple of 2π), and not convergent if s≤0 (apart from the sine series when θ is a multiple of π).
Hint
s>1: domination. 0<s≤1: the corollary with φn=n−s.
Hint
s≤0: the terms do not tend to 0 (see the previous exercise).
Solution
s>1.∣n−scosnθ∣≤n−s and ∑n−s converges: absolute convergence for every θ.
0<s≤1.φn=n−s tends steadily to 0, so both series converge by the corollary, except the cosine series at θ=2kπ, which is ∑n−s and diverges. (They are not absolutely convergent, apart from the trivial sine series at θ=kπ: that is Examples LXXIX, 6, the last exercise below.)
s≤0. By the argument of the previous exercise, ∣cosnθ∣≥21 for infinitely many n (if ∣cosnθ∣<21 then ∣cos2nθ∣>21), and similarly for the sine when θ is not a multiple of π. So the terms, of size n∣s∣∣cosnθ∣, do not tend to 0, and the series cannot converge. For s<0 the terms are even unbounded; since un=sn−sn−1, the partial sums are unbounded too, so the oscillation (if any) is not finite. At θ=2kπ the cosine series is ∑n∣s∣ and diverges to ∞; otherwise Hardy states that the series oscillate infinitely. Unboundedness alone does not show this (the partial sums might tend to +∞ or −∞): one way to finish is to multiply the partial sums of ∑n∣s∣zn, with z=cosθ+isinθ, by 1−z, which isolates a boundary term −N∣s∣zN+1 whose real and imaginary parts take large values of both signs. We omit the details. For s=0 we are back at the previous exercise.
ExerciseRaising the exponent●●○Hardy: Examples LXXIX, 5
If ∑ann−s converges or oscillates finitely, show that ∑ann−t converges for every t>s.
Hint
Write ann−t as a product of two factors suited to Dirichlet’s test.
Solution
ann−t=(ann−s)⋅n−(t−s). The first factor is the general term of a series with bounded partial sums; the second, φn=n−(t−s), tends steadily to 0 because t−s>0. Dirichlet’s test applies. For example, since ∑(−1)n oscillates finitely, ∑(−1)nn−t converges for every t>0.
ExerciseConvergent, but not absolutely●●●Hardy: Examples LXXIX, 6
Let φn tend steadily to 0 with ∑φn divergent. Show that ∑φncosnθ is not absolutely convergent for any θ. Where does the corresponding argument for the sine series break down?
Hint
cos2nθ≤∣cosnθ∣, and cos2nθ=21(1+cos2nθ).
Solution
Suppose ∑φn∣cosnθ∣ converged. Since 0≤cos2nθ≤∣cosnθ∣, so would ∑φncos2nθ=21∑φn+21∑φncos2nθ. If θ is not a multiple of π, then 2θ is not a multiple of 2π and ∑φncos2nθ converges by the corollary; then 21∑φn would converge, which it does not. If θ is a multiple of π, then ∣cosnθ∣=1 and ∑φn∣cosnθ∣=∑φn diverges directly.
For the sine series use sin2nθ=21(1−cos2nθ): the same argument works when θ is not a multiple of π. When it is, sinnθ=0 for all n, the series is identically zero and trivially absolutely convergent: that is the exception.
ExploreSlow decrease still works
Try this.Choose an=sinnθ and φn=1/log(n+1), and show 5000 terms. Compare with φn=1/n.
∑anφn with an=sinnθ,φn=log(n+1)1Partial sums An=a1+⋯+an of the factor an
nAₙ
Partial sums of ∑anφn
nΣ aₖφₖ
Aₙ
band ±K
partial sums of Σ aₙφₙ
room left for the tail: ±2Kφₙ₊₁
Every ∣An∣≤K=1.188 (K=1/∣sin21θ∣). Summation by parts then gives, for every m, ∣amφm+⋯+anφn∣≤2Kφm. After 200 terms the partial sum is 0.677532, and the rest of the series can move it by at most 2Kφ201=0.448.
What to notice
Even though ∑1/log(n+1) diverges far more slowly than anything we can easily sum, the product series converges, just slowly: the purple bar, which bounds the whole remaining tail, shrinks only like 1/logn. Dirichlet’s test asks nothing about how fast φn decreases.
Check yourself
Check yourselfChoose one
Which pair of hypotheses does Dirichlet’s test use for ∑anφn?
Why
Bounded partial sums, not bounded terms: with an=1 the terms are bounded but ∑1/n diverges. The steady decrease of φn is what makes the drops φn−φn+1 positive with a finite total. The first option is too weak in one place and Abel’s test in the other.
Check yourselfTrue or false?
∑n≥1(sinn)/n converges.
Why
True: θ=1 is not a multiple of π, so the partial sums of ∑sinn are bounded by 1/sin21≈2.09, and 1/n tends steadily to 0. (Its sum is (π−1)/2.)
Check yourselfEnter a number
For θ=π/3, what is the bound 1/∣sin21θ∣ for the partial sums of ∑cosnθ?
Why
sin(π/6)=21, so K=2. In fact the partial sums of ∑cos(nπ/3) run 21,0,−1,−23,−1,0,…, which stay inside [−2,2]; the bound is not sharp.
Check yourselfTrue or false?
Since ∑(−1)n−1/n converges, so does ∑(−1)n−1n2n+1.
Why
True, by Abel’s test: (−1)n−1n2n+1=n(−1)n−1⋅(1+n1), and 1+1/n is positive and decreasing (its limit is 1, not 0, which Abel’s test allows).
Hardy’s language
A convergent series remains convergent if we multiply its terms by any sequence of positive and decreasing factors.
— G. H. Hardy, §196
Summary
Summary
Summation by parts: ∑anφn=∑sk(φk−φk+1)+snφn, a discrete integration by parts.
Dirichlet: bounded partial sums of an, times φn decreasing steadily to 0, give a convergent series, with tails at most 2Kφm.
Abel: a convergent ∑an, times positive decreasing factors (any limit), stays convergent.
The partial sums of ∑cosnθ, ∑sinnθ are bounded by 1/∣sin21θ∣, so ∑φncosnθ and ∑φnsinnθ converge (except the cosine series at θ=2kπ), and only conditionally when ∑φn diverges.