Chapter VIII · The convergence of infinite series and infinite integrals

A finer ratio test and the hypergeometric series

§183 30 min

When the ratio of successive terms is 1 − a/n + O(1/n²), the series converges if a > 1 and diverges if a < 1; applied to the hypergeometric series.

Why this matters

d’Alembert’s test fails whenever vn+1/vn→1v_{n+1}/v_n \to 1. But “tends to 11” can happen in very different ways, and the way the ratio approaches 11 carries the information we need. §183 measures it against the ratio for ∑n−s\sum n^{-s}, whose behaviour we now know exactly. The result decides series such as the hypergeometric series of Gauss, which contains the binomial series and many others as special cases, and whose terms are products too complicated for the integral test.

The idea

For un=n−su_n = n^{-s} the ratio of successive terms is

un+1un=(1+1n)−s=1−sn+O ⁣(1n2)\frac{u_{n+1}}{u_n} = \left(1 + \frac1n\right)^{-s} = 1 - \frac{s}{n} + O\!\left(\frac1{n^2}\right)

by Taylor’s theorem. So the ∑n−s\sum n^{-s} are exactly the series whose ratio is “1−s/n1 - s/n, up to a much smaller error”, and among them the dividing line is s=1s = 1. Now take any series whose ratio has the form 1−a/n+O(1/n2)1 - a/n + O(1/n^2). If a>1a > 1, pick ss between 11 and aa: the ratio of our series is eventually smaller than the ratio of the convergent series ∑n−s\sum n^{-s}, so its terms shrink faster, and it converges. If a<1a < 1, the same comparison with a divergent ∑n−s\sum n^{-s} gives divergence.

To read off aa from numbers, look at n(1−vn+1/vn)n(1 - v_{n+1}/v_n), which tends to aa.

Explore Reading aa from the ratios

Try this. Start with α=0.5\alpha = 0.5, β=1.5\beta = 1.5, γ=2.5\gamma = 2.5. Watch n(1−vn+1/vn)n(1 - v_{n+1}/v_n) settle on aa, then lower γ\gamma until the dashed line crosses the red line at 11. Switch to the partial sums on each side of the crossing.

1+αβ1⋅γ+α(α+1)β(β+1)1⋅2⋅γ(γ+1)+⋯1 + \frac{\alpha\beta}{1\cdot\gamma} + \frac{\alpha(\alpha+1)\beta(\beta+1)}{1\cdot 2\cdot\gamma(\gamma+1)} + \cdots
nn(1 − vₙ₊₁/vₙ)
  • n (1−vn+1/vn)n\,(1 - v_{n+1}/v_n)
  • its limit a=γ+1−α−βa = \gamma + 1 - \alpha - \beta
  • the critical value 1
Show
a=γ+1−α−β=1.5a = \gamma + 1 - \alpha - \beta = 1.5. Since a>1a > 1, that is γ>α+β\gamma > \alpha + \beta, the series converges: it is eventually dominated by ∑n−s\sum n^{-s} for any ss between 1 and aa. (Term v300=0.000162v_{300} = 0.000162.)
What to notice

The quantity n(1−vn+1/vn)n(1 - v_{n+1}/v_n) settles on a=γ+1−α−βa = \gamma + 1 - \alpha - \beta, so the series converges exactly when γ>α+β\gamma > \alpha + \beta. For aa a little below 11 the partial sums still look as if they might level off; they do not, just as ∑n−0.9\sum n^{-0.9} does not. At a=1a = 1 exactly the test is silent.

Precise statements

Recall that O(1/n2)O(1/n^2) stands for a quantity whose absolute value is at most K/n2K/n^2 for some constant KK and all large nn (§89).

Theorem A finer ratio test Hardy: §183

Let vn>0v_n > 0 and

vn+1vn=1−an+O ⁣(1n2).(1)\frac{v_{n+1}}{v_n} = 1 - \frac{a}{n} + O\!\left(\frac{1}{n^2}\right). \tag{1}

Then ∑vn\sum v_n converges if a>1a > 1 and diverges if a<1a < 1. Moreover, if a>0a > 0 then vn→0v_n \to 0.

The case a=1a = 1 is left by Hardy to Chapter IX (Examples XC, 5), where comparison with ∑1/(nlog⁡n)\sum 1/(n\log n) shows that the series then diverges.

The hypergeometric series is

1+αβ1⋅γ+α(α+1)β(β+1)1⋅2⋅γ(γ+1)+α(α+1)(α+2)β(β+1)(β+2)1⋅2⋅3⋅γ(γ+1)(γ+2)+⋯ ,(2)1 + \frac{\alpha\beta}{1\cdot\gamma} + \frac{\alpha(\alpha+1)\beta(\beta+1)}{1\cdot2\cdot\gamma(\gamma+1)} + \frac{\alpha(\alpha+1)(\alpha+2)\beta(\beta+1)(\beta+2)}{1\cdot2\cdot3\cdot\gamma(\gamma+1)(\gamma+2)} + \cdots, \tag{2}

where α\alpha, β\beta, γ\gamma are real and none of them is zero or a negative integer.

Corollary Convergence of the hypergeometric series Hardy: §183

The series (2) converges if γ>α+β\gamma > \alpha + \beta and diverges if γ<α+β\gamma < \alpha + \beta. Its terms tend to 00 if γ>α+β−1\gamma > \alpha + \beta - 1. In particular the series

1+m1+m(m+1)1⋅2+⋯1 + \frac{m}{1} + \frac{m(m+1)}{1\cdot2} + \cdots

converges if m<0m < 0 and diverges if m>0m > 0.

The last series is (2) with α=m\alpha = m, β=γ=1\beta = \gamma = 1; it is the binomial series of (1−x)−m(1-x)^{-m} at x=1x = 1.

Proofs

Proof The finer ratio test
Strategy

Compare the ratio of ∑vn\sum v_n with the ratio of ∑n−s\sum n^{-s} for a well-chosen ss, and use the comparison of ratios from §175: if vn+1/vn≤un+1/unv_{n+1}/v_n \le u_{n+1}/u_n for large nn and ∑un\sum u_n converges, so does ∑vn\sum v_n.

The ratio for n−sn^{-s}

By Taylor’s theorem with remainder (§151), applied to (1+h)−s(1 + h)^{-s} with h=1/nh = 1/n,

(1+1n)−s=1−sn+s(s+1)2n2(1+θn)−s−2,0<θ<1,\left(1 + \frac1n\right)^{-s} = 1 - \frac sn + \frac{s(s+1)}{2n^2}\left(1 + \frac\theta n\right)^{-s-2}, \qquad 0 < \theta < 1,

and the last term is O(1/n2)O(1/n^2) for s>0s > 0. So un+1/un=1−s/n+O(1/n2)u_{n+1}/u_n = 1 - s/n + O(1/n^2).

a>1a > 1: choose 1<s<a1 < s < a

Subtracting,

un+1un−vn+1vn=a−sn+O ⁣(1n2).\frac{u_{n+1}}{u_n} - \frac{v_{n+1}}{v_n} = \frac{a - s}{n} + O\!\left(\frac1{n^2}\right).

Since a−s>0a - s > 0, the first term beats the error for large nn: the error is at most K/n2K/n^2, which is less than (a−s)/n(a-s)/n once n>K/(a−s)n > K/(a-s). So vn+1/vn<un+1/unv_{n+1}/v_n < u_{n+1}/u_n for large nn.

Conclude convergence

∑n−s\sum n^{-s} converges because s>1s > 1, so by the comparison of ratios ∑vn\sum v_n converges.

a<1a < 1: choose a<s<1a < s < 1

Now a−s<0a - s < 0 and the same computation gives vn+1/vn>un+1/unv_{n+1}/v_n > u_{n+1}/u_n for large nn. Since ∑n−s\sum n^{-s} diverges for s<1s < 1, the divergence half of the comparison of ratios shows that ∑vn\sum v_n diverges. (If a≤0a \le 0, take any ss with 0<s<10 < s < 1 and s>as > a: the same argument works.)

vn→0v_n \to 0 when a>0a > 0

Choose 0<s<a0 < s < a. As above vn+1/vn<un+1/unv_{n+1}/v_n < u_{n+1}/u_n for n≥n0n \ge n_0, and the telescoping argument of §175 gives vn≤Kn−sv_n \le K n^{-s} for n≥n0n \ge n_0. Hence vn→0v_n \to 0.

Proof The hypergeometric series
Strategy

Compute the ratio of successive terms, expand it in powers of 1/n1/n, and read off aa.

The terms are eventually of one sign

The factors α+n\alpha + n, β+n\beta + n, γ+n\gamma + n are positive once nn exceeds −α-\alpha, −β-\beta and −γ-\gamma, and none of them is ever zero. So from some point on all terms have the same sign, and changing the sign of the whole series if necessary we may apply the test to positive terms. (Finitely many early terms do not matter.)

The ratio

Writing vnv_n for the term containing nn factors in each product,

vn+1vn=(α+n)(β+n)(n+1)(γ+n).\frac{v_{n+1}}{v_n} = \frac{(\alpha + n)(\beta + n)}{(n + 1)(\gamma + n)} .

Expand

The numerator is n2+(α+β)n+αβn^2 + (\alpha + \beta)n + \alpha\beta and the denominator n2+(γ+1)n+γn^2 + (\gamma + 1)n + \gamma. Their difference is (γ+1−α−β)n+(γ−αβ)(\gamma + 1 - \alpha - \beta)n + (\gamma - \alpha\beta), so

vn+1vn=1−(γ+1−α−β)n+(γ−αβ)(n+1)(γ+n)=1−γ+1−α−βn+O ⁣(1n2).\frac{v_{n+1}}{v_n} = 1 - \frac{(\gamma + 1 - \alpha - \beta)n + (\gamma - \alpha\beta)}{(n+1)(\gamma + n)} = 1 - \frac{\gamma + 1 - \alpha - \beta}{n} + O\!\left(\frac1{n^2}\right).

Apply the test

Here a=γ+1−α−βa = \gamma + 1 - \alpha - \beta. The series converges if a>1a > 1, that is γ>α+β\gamma > \alpha + \beta, and diverges if γ<α+β\gamma < \alpha + \beta; its terms tend to 00 if a>0a > 0, that is γ>α+β−1\gamma > \alpha + \beta - 1. For the special series, a=1+1−m−1=1−ma = 1 + 1 - m - 1 = 1 - m, which exceeds 11 exactly when m<0m < 0.

Worked examples and exercises

Exercise Products of odd and even numbers ●●○

Let cn=1⋅3⋅5⋯(2n−1)2⋅4⋅6⋯(2n)c_n = \dfrac{1\cdot3\cdot5\cdots(2n-1)}{2\cdot4\cdot6\cdots(2n)}. Show that ∑cn\sum c_n diverges, that ∑cn3\sum c_n^3 converges, and that ∑cn2\sum c_n^2 is a borderline case of the test.

Hint
cn+1/cn=(2n+1)/(2n+2)c_{n+1}/c_n = (2n+1)/(2n+2). Write it as 1−12n+21 - \dfrac{1}{2n+2} and expand.
Hint
The ratio for cnpc_n^p is the pp-th power of the ratio for cnc_n, and (1−x)p=1−px+O(x2)(1 - x)^p = 1 - px + O(x^2).

Solution

cn+1cn=2n+12n+2=1−12(n+1)=1−12n+O ⁣(1n2)\dfrac{c_{n+1}}{c_n} = \dfrac{2n+1}{2n+2} = 1 - \dfrac{1}{2(n+1)} = 1 - \dfrac{1}{2n} + O\!\left(\dfrac{1}{n^2}\right), so a=12<1a = \tfrac12 < 1 and ∑cn\sum c_n diverges (though cn→0c_n \to 0, since a>0a > 0).

For cnpc_n^p the ratio is (1−12n+O(n−2))p=1−p2n+O ⁣(1n2)\big(1 - \tfrac{1}{2n} + O(n^{-2})\big)^p = 1 - \dfrac{p}{2n} + O\!\left(\dfrac1{n^2}\right), so a=p/2a = p/2. For p=3p = 3, a=32>1a = \tfrac32 > 1: convergent. For p=2p = 2, a=1a = 1: the test is silent; Chapter IX shows divergence. (In fact cn∼1/πnc_n \sim 1/\sqrt{\pi n}, so ∑cnp\sum c_n^p converges exactly when p>2p > 2.)

Exercise Factorials over rising products ●●○

Let α>0\alpha > 0. Show that ∑n≥1n!(α+1)(α+2)⋯(α+n)\displaystyle\sum_{n\ge1} \frac{n!}{(\alpha+1)(\alpha+2)\cdots(\alpha+n)} converges if α>1\alpha > 1 and diverges if α<1\alpha < 1. What happens when α=1\alpha = 1?

Hint
The ratio of successive terms is (n+1)/(α+n+1)(n+1)/(\alpha + n + 1).

Solution

vn+1vn=n+1n+1+α=1−αn+1+α=1−αn+O ⁣(1n2)\dfrac{v_{n+1}}{v_n} = \dfrac{n+1}{n+1+\alpha} = 1 - \dfrac{\alpha}{n+1+\alpha} = 1 - \dfrac{\alpha}{n} + O\!\left(\dfrac{1}{n^2}\right), since αn−αn+1+α=α(1+α)n(n+1+α)\dfrac{\alpha}{n} - \dfrac{\alpha}{n+1+\alpha} = \dfrac{\alpha(1+\alpha)}{n(n+1+\alpha)}. So a=αa = \alpha: convergence for α>1\alpha > 1, divergence for α<1\alpha < 1. When α=1\alpha = 1 the term is n!/(n+1)!=1/(n+1)n!/(n+1)! = 1/(n+1), and the series is the harmonic series without its first term: divergent, as Chapter IX’s result for a=1a = 1 predicts.

Exercise Two familiar series ●○○ Hardy: §183

For vn=1/n2v_n = 1/n^2 and vn=1/(n(n+1))v_n = 1/(n(n+1)), compute aa in (1), and check that the test gives the right answer.

Solution

(n+1)−2n−2=(1+1n)−2=1−2n+O ⁣(1n2)\dfrac{(n+1)^{-2}}{n^{-2}} = \left(1 + \dfrac1n\right)^{-2} = 1 - \dfrac2n + O\!\left(\dfrac1{n^2}\right), so a=2a = 2. And 1/((n+1)(n+2))1/(n(n+1))=nn+2=1−2n+2=1−2n+O ⁣(1n2)\dfrac{1/((n+1)(n+2))}{1/(n(n+1))} = \dfrac{n}{n+2} = 1 - \dfrac{2}{n+2} = 1 - \dfrac2n + O\!\left(\dfrac1{n^2}\right), so again a=2a = 2. Both converge. The finer test sees the difference between 1−2/n1 - 2/n and 1−1/n1 - 1/n (the harmonic series), which d’Alembert’s test, looking only at the limit 11, cannot.

Exercise The binomial series at x=1x = 1 ●●○ Hardy: §183

For which real mm (not 00 or a negative integer) does 1+m+m(m+1)2!+m(m+1)(m+2)3!+⋯1 + m + \dfrac{m(m+1)}{2!} + \dfrac{m(m+1)(m+2)}{3!} + \cdots converge? Check your answer for m=−12m = -\tfrac12 by writing out the first terms.

Solution

This is (2) with α=m\alpha = m, β=γ=1\beta = \gamma = 1, so a=1−ma = 1 - m: convergence for m<0m < 0, divergence for m>0m > 0 (for m>0m > 0 every term is at least 11 anyway). For m=−12m = -\tfrac12 the terms are 1,−12,−18,−116,−5128,…1, -\tfrac12, -\tfrac18, -\tfrac1{16}, -\tfrac{5}{128}, \dots: after the first term they all have the same sign, as the proof requires, and their size decreases like n−3/2n^{-3/2}. The sum is (1−1)1/2=0(1 - 1)^{1/2} = 0, which is the binomial theorem at the edge of its range (§222).

Check yourself

Check yourself Choose one

A positive series has vn+1/vn=1−2n+O ⁣(1n2)v_{n+1}/v_n = 1 - \dfrac{2}{n} + O\!\left(\dfrac{1}{n^2}\right). What follows?

Why

Here a=2>1a = 2 > 1, so the series converges; it behaves roughly like ∑n−2\sum n^{-2}. d’Alembert’s test does fail (the ratio tends to 11), which is exactly why the finer test is needed.

Check yourself True or false?

When a=1a = 1 in (1), the test of §183 shows that the series diverges.

Why

False, as far as §183 goes: the argument needs an ss strictly between aa and 11, which does not exist when a=1a = 1. The divergence in that case is proved in Ch. IX by comparing with ∑1/(nlog⁡n)\sum 1/(n \log n), a series that diverges more slowly than any ∑n−s\sum n^{-s} with s<1s < 1.

Check yourself Enter a number

For the hypergeometric series with α=1\alpha = 1, β=2\beta = 2, γ=3.5\gamma = 3.5, what is aa in (1)?

Why

a=γ+1−α−β=3.5+1−1−2=1.5>1a = \gamma + 1 - \alpha - \beta = 3.5 + 1 - 1 - 2 = 1.5 > 1, so the series converges. Equivalently, γ=3.5>α+β=3\gamma = 3.5 > \alpha + \beta = 3.

Check yourself True or false?

If γ=α+β−12\gamma = \alpha + \beta - \tfrac12, the terms of the hypergeometric series tend to 00 but the series diverges.

Why

True. Then a=12a = \tfrac12: since 0<a<10 < a < 1, the terms tend to 00 (they behave like n−1/2n^{-1/2}) and the series diverges.

Hardy’s language

Summary

Summary
  • The ratio for ∑n−s\sum n^{-s} is 1−s/n+O(1/n2)1 - s/n + O(1/n^2). Comparing ratios with it gives a test that sees how the ratio tends to 11.
  • If vn+1/vn=1−a/n+O(1/n2)v_{n+1}/v_n = 1 - a/n + O(1/n^2): convergence for a>1a > 1, divergence for a<1a < 1; a=1a = 1 needs Ch. IX (divergence). If a>0a > 0 the terms tend to 00.
  • In practice: compute n(1−vn+1/vn)n(1 - v_{n+1}/v_n) and find its limit aa.
  • The hypergeometric series has a=γ+1−α−βa = \gamma + 1 - \alpha - \beta: it converges exactly when γ>α+β\gamma > \alpha + \beta.