Chapter VIII · The convergence of infinite series and infinite integrals

Ratio tests and d'Alembert's test

§175 30 min

Comparing successive terms instead of the terms themselves, d'Alembert's ratio test, and why it is less general than Cauchy's root test.

Why this matters

The root test of the last lesson needs vn1/nv_n^{1/n}, which is awkward when vnv_n contains factorials or a product of many factors. In §175 Hardy compares successive terms instead. The ratio vn+1/vnv_{n+1}/v_n of a complicated term is often very simple: for vn=n!/nnv_n = n!/n^n it is (n/(n+1))n(n/(n+1))^n. The resulting test, d’Alembert’s, is the one most people reach for first. This lesson is about how it works, why it works, and where it fails.

The idea

A geometric series is the series in which every term is the same multiple rr of the one before. If the ratio of successive terms of ∑vn\sum v_n is eventually at most rr, with r<1r < 1, then from some point on the terms shrink at least as fast as a geometric series, and the series converges for the same reason the geometric series does. On a logarithmic scale this is easy to see: a geometric series is a straight line going down, and a series whose ratios tend to rr eventually runs parallel to that line.

Explore Terms that shrink geometrically

Try this. Plot the terms of ∑nkrn\sum n^k r^n with the logarithmic scale on, for k=2k = 2 and r=0.7r = 0.7. Then raise kk to 55 and show 100 terms.

∑n=1∞nkrn\sum_{n=1}^{\infty} n^k r^n
nuₙ (log scale)
  • uₙ for nkrnn^k r^n
Plot
After 100 terms s100=44.07407407s_{100} = 44.07407407. The series converges.
What to notice

At first the factor nkn^k makes the terms grow, but the plot always bends over into a straight line of slope log⁡10r\log_{10} r. That is because vn+1/vn=(1+1/n)kr→rv_{n+1}/v_n = (1 + 1/n)^k r \to r: eventually every step multiplies the term by nearly rr. A large kk only delays this.

The ratio test only looks at neighbouring terms, and that is its weakness. A series can converge perfectly well while its neighbouring terms keep jumping up and down.

Explore Where the ratio test is blind

Try this. Compare the ratio test and the root test on a+b+a2+b2+a3+⋯a + b + a^2 + b^2 + a^3 + \cdots with a=14a = \tfrac14, b=12b = \tfrac12. Then try both tests on 1/n!1/n!, n!/nnn!/n^n and 1/n21/n^2.

∑n=1∞vn,vn=v2m−1=am, v2m=bm\sum_{n=1}^{\infty} v_n,\quad v_n = v_{2m-1} = a^m,\ v_{2m} = b^m
nvₙ₊₁/vₙ
  • vn+1/vnv_{n+1}/v_n (below 1)
  • the critical value 1 (points at or above it)
Test
At n=49n = 49: vn+1/vn=3.3554×107v_{n+1}/v_n = 3.3554 × 10⁷. Hollow points are off the top of the plot. Test inconclusive or not applicable. The ratio alternates between (b/a)m→∞(b/a)^m \to \infty and a(a/b)m→0a(a/b)^m \to 0. It is neither eventually ≤r<1\le r < 1 nor eventually ≥1\ge 1: d’Alembert’s test fails completely (Ex. LXVIII. 9), although the series converges.
What to notice

For the interleaved series the ratios jump between huge and tiny values (the hollow points are off the scale), so d’Alembert’s test cannot be applied, while every root stays below b≈0.71\sqrt b \approx 0.71 and Cauchy’s test proves convergence. For 1/n!1/n! and n!/nnn!/n^n the ratio settles quickly and is easy to compute. For 1/n21/n^2 both tests sit at 11 and say nothing.

Precise statements

In this lesson all terms are strictly positive, since we divide by them.

Theorem Comparison of ratios Hardy: §175, 4 and 5

Let un>0u_n > 0 and vn>0v_n > 0.

  1. If vn+1vn≤un+1un\dfrac{v_{n+1}}{v_n} \le \dfrac{u_{n+1}}{u_n} for all sufficiently large nn, and ∑un\sum u_n converges, then ∑vn\sum v_n converges.
  2. If vn+1vn≥un+1un\dfrac{v_{n+1}}{v_n} \ge \dfrac{u_{n+1}}{u_n} for all sufficiently large nn, and ∑un\sum u_n diverges, then ∑vn\sum v_n diverges.

Taking un=rnu_n = r^n in part 1 gives the test itself.

Theorem d’Alembert’s test Hardy: §175, 6

If vn+1/vn≤rv_{n+1}/v_n \le r for all sufficiently large nn, where r<1r < 1, then ∑vn\sum v_n converges. If vn+1/vn≥1v_{n+1}/v_n \ge 1 for all sufficiently large nn, then ∑vn\sum v_n diverges.

Most often the ratio, or the root, has a limit, and then the tests take their familiar form.

Corollary The limit forms Hardy: §175, 7

If vn1/nv_n^{1/n} or vn+1/vnv_{n+1}/v_n tends to a limit ll as n→∞n \to \infty, then ∑vn\sum v_n converges if l<1l < 1 and diverges if l>1l > 1. If l=1l = 1, the test gives no information.

The case l=1l = 1 really is undecided: ∑1/n\sum 1/n diverges, ∑1/n2\sum 1/n^2 converges, and for both of them the ratio and the root tend to 11. The tests also fail when the ratio or root stays below 11 but comes arbitrarily close to it, and the ratio test fails when the ratio is sometimes above and sometimes below 11, as in the interleaved series above.

Proofs

Proof Comparison of ratios
Strategy

Write vnv_n as vn0v_{n_0} times a chain of ratios, compare the chain link by link with the chain for unu_n, and conclude vn≤Kunv_n \le Ku_n with a constant KK. Then Theorem C finishes the proof.

Telescope the ratios

For n>n0n > n_0,

vn=vnvn−1⋅vn−1vn−2⋯vn0+1vn0  vn0.v_n = \frac{v_n}{v_{n-1}}\cdot\frac{v_{n-1}}{v_{n-2}}\cdots\frac{v_{n_0+1}}{v_{n_0}}\; v_{n_0}.

Compare factor by factor

Each ratio vν+1/vνv_{\nu+1}/v_\nu with ν≥n0\nu \ge n_0 is at most uν+1/uνu_{\nu+1}/u_\nu, and all factors are positive, so

vn≤unun−1⋯un0+1un0  vn0=vn0un0 un.v_n \le \frac{u_n}{u_{n-1}}\cdots\frac{u_{n_0+1}}{u_{n_0}}\; v_{n_0} = \frac{v_{n_0}}{u_{n_0}}\, u_n .

A constant KK

K=vn0/un0K = v_{n_0}/u_{n_0} does not depend on nn. So vn≤Kunv_n \le K u_n for n≥n0n \ge n_0, and Theorem C gives convergence of ∑vn\sum v_n.

The divergence half

With the inequalities reversed the same chain gives vn≥Kunv_n \ge K u_n, with K>0K > 0, and Theorem C gives divergence.

Proof d’Alembert’s test and its limit form
Strategy

Apply the comparison of ratios with the geometric series, then show that a limit below 11 (or above 11) produces the hypothesis needed.

Convergence

Take un=rnu_n = r^n, so un+1/un=ru_{n+1}/u_n = r. If vn+1/vn≤rv_{n+1}/v_n \le r for large nn and r<1r < 1, part 1 of the comparison applies, since ∑rn\sum r^n converges.

Divergence

If vn+1/vn≥1v_{n+1}/v_n \ge 1 for n≥n0n \ge n_0, then vn≥vn0>0v_n \ge v_{n_0} > 0 for all n≥n0n \ge n_0. The terms do not tend to 00, so the series diverges.

A limit l<1l < 1

Choose rr with l<r<1l < r < 1, for example r=12(1+l)r = \tfrac12(1 + l). Taking δ=r−l\delta = r - l in the definition of a limit, vn+1/vn<rv_{n+1}/v_n < r for all large nn, and the series converges. The root version is the same with Cauchy’s test in place of d’Alembert’s.

A limit l>1l > 1

Now vn+1/vn>1v_{n+1}/v_n > 1 for all large nn (take δ=l−1\delta = l - 1), so the terms eventually increase and cannot tend to 00. For the root, vn1/n>1v_n^{1/n} > 1 for large nn means vn>1v_n > 1, with the same conclusion. Hardy leaves this half to the reader.

Hardy remarks that d’Alembert’s test is theoretically less general than Cauchy’s: whenever the ratio test proves convergence, so does the root test. The proof is short.

Proof Ratio test succeeds ⟹ root test succeeds
Strategy

The ratio hypothesis gives vn≤Krnv_n \le Kr^n. Taking nn-th roots turns the constant KK into K1/nK^{1/n}, which tends to 11 and so does no harm.

From ratios to a bound

If vn+1/vn≤r<1v_{n+1}/v_n \le r < 1 for n≥n0n \ge n_0, the proof above gives vn≤Krnv_n \le K r^n with K=vn0/rn0K = v_{n_0}/r^{n_0}.

Take roots

Then vn1/n≤K1/nrv_n^{1/n} \le K^{1/n} r. Since K1/n→1K^{1/n} \to 1 (Examples XXVII, 10), for any r′r' with r<r′<1r < r' < 1 we get vn1/n≤r′v_n^{1/n} \le r' for all large nn, which is Cauchy’s hypothesis.

The converse fails

The interleaved series a+b+a2+b2+⋯a + b + a^2 + b^2 + \cdots satisfies Cauchy’s hypothesis with r′=br' = \sqrt b, but its ratios are not eventually below any r<1r < 1 (Exercise 4 below).

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXVIII, 1

Let kk be a positive integer and r>0r > 0. Use d’Alembert’s test to show that ∑nkrn\sum n^k r^n converges if r<1r < 1 and diverges if r≥1r \ge 1. Check that Cauchy’s test gives the same.

Hint
Compute vn+1/vnv_{n+1}/v_n and let n→∞n \to \infty.
Hint
For the root test you need lim⁡n1/n=1\lim n^{1/n} = 1 (Examples XXVII, 11).

Solution
vn+1vn=(n+1n)kr→r.\frac{v_{n+1}}{v_n} = \left(\frac{n+1}{n}\right)^k r \to r .

So the series converges if r<1r < 1 and diverges if r>1r > 1. When r=1r = 1 the test fails, but then vn=nk≥1v_n = n^k \ge 1, so the terms do not tend to 00 and the series diverges. For the root test, vn1/n=(n1/n)k r→rv_n^{1/n} = (n^{1/n})^k\, r \to r since n1/n→1n^{1/n} \to 1, with the same conclusions. In particular nkrn→0n^k r^n \to 0 when r<1r < 1, however large kk is.

Exercise ●●○ Hardy: Examples LXVIII, 2, 3 and 6

Let PP and QQ be polynomials of degrees kk and ll with positive leading coefficients, and r>0r > 0. Discuss the convergence of ∑P(n)Q(n) rn\sum \dfrac{P(n)}{Q(n)}\, r^n (start the sum beyond the largest root of QQ).

Hint
P(n)/Q(n)∼(A/α) nk−lP(n)/Q(n) \sim (A/\alpha)\, n^{k-l}, where AA and α\alpha are the leading coefficients. Use Theorem D.
Hint
When r=1r = 1 the ratio test fails. Compare with ∑nk−l\sum n^{k-l} and use the last lesson.

Solution

By Theorem D, the series behaves like ∑nk−lrn\sum n^{k-l} r^n, since the ratio of the terms tends to A/α≠0A/\alpha \ne 0.

If r≠1r \ne 1, d’Alembert’s test applies as in the previous exercise (the exponent k−lk - l may now be negative, which changes nothing): (n+1n)k−lr→r(\tfrac{n+1}{n})^{k-l} r \to r. So the series converges for r<1r < 1 and diverges for r>1r > 1.

If r=1r = 1 the series behaves like ∑n−(l−k)\sum n^{-(l-k)}, which converges exactly when l−k≥2l - k \ge 2, that is l>k+1l > k + 1, and diverges when l≤k+1l \le k + 1 (Examples LXVIII, 5, and the divergence of ∑1/n\sum 1/n; for exponents that are not integers see §181).

Exercise ●○○ Hardy: Examples LXVIII, 4

Show that Cauchy’s and d’Alembert’s tests both fail for ∑1n(n+1)\sum \dfrac{1}{n(n+1)}, although the series converges.

Solution

vn+1vn=nn+2→1\dfrac{v_{n+1}}{v_n} = \dfrac{n}{n+2} \to 1, and vn1/n=(n1/n(n+1)1/n)−1→1v_n^{1/n} = \big(n^{1/n}(n+1)^{1/n}\big)^{-1} \to 1, since n1/n→1n^{1/n} \to 1 and (n+1)1/n→1(n+1)^{1/n} \to 1. Both limits equal 11, so both tests are silent, yet the series telescopes to the sum 11. The geometric series is too coarse a yardstick for series whose terms decrease like a power of nn.

Exercise ●●○ Hardy: Examples LXVIII, 9

Let 0<a<b<10 < a < b < 1. Show that a+b+a2+b2+a3+b3+⋯a + b + a^2 + b^2 + a^3 + b^3 + \cdots converges, that Cauchy’s test proves it, and that d’Alembert’s test fails.

Hint
Write v2m−1=amv_{2m-1} = a^m, v2m=bmv_{2m} = b^m, and compute the two kinds of ratio.

Solution

The series is the sum of two convergent geometric series, so it converges (to a/(1−a)+b/(1−b)a/(1-a) + b/(1-b)).

Ratios. v2m/v2m−1=(b/a)m→∞v_{2m}/v_{2m-1} = (b/a)^m \to \infty and v2m+1/v2m=a(a/b)m→0v_{2m+1}/v_{2m} = a(a/b)^m \to 0. The ratios are neither eventually ≤r<1\le r < 1 nor eventually ≥1\ge 1, so d’Alembert’s test says nothing.

Roots. v2m1/(2m)=bv_{2m}^{1/(2m)} = \sqrt b, and v2m−11/(2m−1)=am/(2m−1)≤a<bv_{2m-1}^{1/(2m-1)} = a^{m/(2m-1)} \le \sqrt a < \sqrt b because m/(2m−1)≥12m/(2m-1) \ge \tfrac12 and a<1a < 1. So vn1/n≤b<1v_n^{1/n} \le \sqrt b < 1 for every nn, and Cauchy’s test proves convergence.

Exercise ●●○ Hardy: Examples LXVIII, 10 and 11

Show that ∑rn/n!\sum r^n/n! and ∑rn/nn\sum r^n/n^n converge for every r>0r > 0, that ∑n! rn\sum n!\,r^n and ∑nnrn\sum n^n r^n diverge for every r>0r > 0, and that ∑(nrn+1)n\sum \left(\dfrac{nr}{n+1}\right)^n converges if r<1r < 1 and diverges if r≥1r \ge 1.

Hint
Use the ratio test for the factorials and the root test for the nn-th powers.
Hint
For the last series at r=1r = 1, recall (1+1/n)n→e(1 + 1/n)^n \to e (§73).

Solution

rn+1/(n+1)!rn/n!=rn+1→0\dfrac{r^{n+1}/(n+1)!}{r^n/n!} = \dfrac{r}{n+1} \to 0, so ∑rn/n!\sum r^n/n! converges; and (rn/nn)1/n=r/n→0(r^n/n^n)^{1/n} = r/n \to 0, so ∑rn/nn\sum r^n/n^n converges.

(n+1)! rn+1n! rn=(n+1)r→∞\dfrac{(n+1)!\,r^{n+1}}{n!\,r^n} = (n+1)r \to \infty and (nnrn)1/n=nr→∞(n^n r^n)^{1/n} = nr \to \infty, so the other two diverge (their terms grow without bound).

For the last series, vn1/n=nr/(n+1)→rv_n^{1/n} = nr/(n+1) \to r: convergence for r<1r < 1, divergence for r>1r > 1. At r=1r = 1 the test fails, but vn=(1+1/n)−n→1/e≠0v_n = (1 + 1/n)^{-n} \to 1/e \ne 0, so the series diverges.

Check yourself

Check yourself True or false?

If vn>0v_n > 0 and vn+1/vn<1v_{n+1}/v_n < 1 for every nn, then ∑vn\sum v_n converges.

Why

False. For vn=1/nv_n = 1/n every ratio n/(n+1)n/(n+1) is less than 11, but the series diverges. The ratios must stay below a fixed r<1r < 1; here they approach 11. A strictly decreasing sequence of terms is not enough.

Check yourself Choose one

For a series of positive terms, vn+1/vn→1v_{n+1}/v_n \to 1. What can you conclude?

Why

A limit 11 leaves the question open. The terms may still tend to 00 (as 1/n1/n and 1/n21/n^2 do), so the second option is wrong too. The root test does not help: if the ratio tends to 11 then so does the root (Hardy proves this in Ch. IX), so it fails as well. Finer tests come in §§178–183.

Check yourself Enter a number

For vn=n5/3nv_n = n^5/3^n, what is lim⁡vn+1/vn\lim v_{n+1}/v_n?

Why

vn+1vn=(1+1n)513→13\dfrac{v_{n+1}}{v_n} = \left(1 + \dfrac1n\right)^5 \dfrac13 \to \dfrac13. The series converges, although its first terms grow (up to about n=5n = 5).

Check yourself True or false?

There are series of positive terms whose convergence d’Alembert’s test proves but Cauchy’s test cannot.

Why

False. If vn+1/vn≤r<1v_{n+1}/v_n \le r < 1 for large nn then vn≤Krnv_n \le Kr^n, so vn1/n≤K1/nrv_n^{1/n} \le K^{1/n} r, which is below any r′∈(r,1)r' \in (r, 1) for large nn. The root test is the stronger one in theory; the ratio test is often easier in practice.

Check yourself Choose one

Does ∑n!/10n\sum n!/10^n converge?

Why

vn+1/vn=(n+1)/10→∞v_{n+1}/v_n = (n+1)/10 \to \infty. From n=10n = 10 on, each term is at least as large as the one before, so the terms do not tend to 00. Factorials eventually beat every geometric growth.

Hardy’s language

Ratio tests are useless for ‘irregular’ series, such as 0 + ½ + 0 + ¼ + 0 + ⅛ + ….

— G. H. Hardy, §175

Summary

Summary
  • If un,vn>0u_n, v_n > 0 and vn+1/vn≤un+1/unv_{n+1}/v_n \le u_{n+1}/u_n for large nn, then vn≤Kunv_n \le Ku_n: ratios can be compared instead of terms.
  • d’Alembert’s test: vn+1/vn≤r<1v_{n+1}/v_n \le r < 1 for large nn gives convergence; vn+1/vn≥1v_{n+1}/v_n \ge 1 for large nn gives divergence.
  • Limit forms: a ratio or root limit below 11 means convergence, above 11 divergence, equal to 11 means no information.
  • The root test is more general; the ratio test is easier for factorials and products.
  • Both tests use the geometric series as a yardstick, which is far too coarse for terms like 1/n21/n^2: finer tests follow.