Chapter VIII
The Convergence of Infinite Series and Infinite Integrals

165. IN Ch. IV we explained what was meant by saying that an infinite series is convergent, divergent, or oscillatory, and illustrated our definitions by a few simple examples, mainly derived from the geometrical series

1 + x + x2 +
and other series closely connected with it. In this chapter we shall pursue the subject in a more systematic manner, and prove a number of theorems which enable us to determine when the simplest series which commonly occur in analysis are convergent.

We shall often use the notation

um + um+1 + + un = mnφ(ν),
and write 0u n, or simply un, for the infinite series u0 + u1 + u2 + .86

166. Series of Positive Terms. The theory of the convergence of series is comparatively simple when all the terms of the series considered are positive.87 We shall consider such series first, not only because they are the easiest to deal with, but also because the discussion of the convergence of a series containing negative or complex terms can often be made to depend upon a similar discussion of a series of positive terms only.

When we are discussing the convergence or divergence of a series we may disregard any finite number of terms. Thus, when a series contains a finite number only of negative or complex terms, we may omit them and apply the theorems which follow to the remainder.

167. It will be well to recall the following fundamental theorems established in § 77.

A. A series of positive terms must be convergent or diverge to , and cannot oscillate.

B. The necessary and sufficient condition that un should be convergent is that there should be a number K such that

u0 + u1 + + un < K
for all values of n.

C. The comparison theorem. If un is convergent, and vn un for all values of n, then vn is convergent, and vn un. More generally, if vn Kun, where K is a constant, then vn is convergent and vn K un. And if un is divergent, and vn Kun, then vn is divergent.88

Moreover, in inferring the convergence or divergence of  vn by means of one of these tests, it is sufficient to know that the test is satisfied for sufficiently large values of n, i.e. for all values of n greater than a definite value n0. But of course the conclusion that vn K un does not necessarily hold in this case.

A particularly useful case of this theorem is

D. If un is convergent divergent and un/vn tends to a limit other than zero as n , then vn is convergent divergent.

168. First applications of these tests. The one important fact which we know at present, as regards the convergence of any special class of series, is that rn is convergent if r < 1 and divergent if r 1.89 It is therefore natural to try to apply Theorem C, taking un = rn. We at once find

1. The series  vn is convergent if vn Krn, where r < 1, for all sufficiently large values of n.

When K = 1, this condition may be written in the form vn1/n r. Hence we obtain what is known as Cauchy’s test for the convergence of a series of positive terms; viz.

2. The series  vn is convergent if vn1/n r, where r < 1, for all sufficiently large values of n.

There is a corresponding test for divergence, viz.

2a. The series  vn is divergent if vn1/n 1 for an infinity of values of n.

This hardly requires proof, for vn1/n 1 involves vn 1. The two theorems 2 and 2a are of very wide application, but for some purposes it is more convenient to use a different test of convergence, viz.

3. The series  vn is convergent if vn+1/vn r, r < 1, for all sufficiently large values of n.

To prove this we observe that if vn+1/vn r when n n0 then

vn = vn vn1vn1 vn2vn0+1 vn0 vn0 vn0 rn0 rn;
and the result follows by comparison with the convergent series  rn. This test is known as d’Alembert’s test. We shall see later that it is less general, theoretically, than Cauchy’s, in that Cauchy’s test can be applied whenever d’Alembert’s can, and sometimes when the latter cannot. Moreover the test for divergence which corresponds to d’Alembert’s test for convergence is much less general than the test given by Theorem 2a. It is true, as the reader will easily prove for himself, that if vn+1/vn r 1 for all values of n, or all sufficiently large values, then vn is divergent. But it is not true (see Ex. LXVII. 9) that this is so if only vn+1/vn r 1 for an infinity of values of n, whereas in Theorem 2a our test had only to be satisfied for such an infinity of values. None the less d’Alembert’s test is very useful in practice, because when vn is a complicated function vn+1/vn is often much less complicated and so easier to work with.

In the simplest cases which occur in analysis it often happens that vn+1/vn or vn1/n tends to a limit as n .90 When this limit is less than 1, it is evident that the conditions of Theorems 2 or 3 above are satisfied. Thus

4. If vn1/n or vn+1/vn tends to a limit less than unity as n , then the series  vn is convergent.

It is almost obvious that if either function tend to a limit greater than unity, then vn is divergent. We leave the formal proof of this as an exercise to the reader. But when vn1/n or vn+1/vn tends to 1 these tests generally fail completely, and they fail also when vn1/n or vn+1/vn oscillates in such a way that, while always less than 1, it assumes for an infinity of values of n values approaching indefinitely near to 1. And the tests which involve vn+1/vn fail even when that ratio oscillates so as to be sometimes less than and sometimes greater than 1. When vn1/n behaves in this way Theorem 2a is sufficient to prove the divergence of the series. But it is clear that there is a wide margin of cases in which some more subtle tests will be needed.

Examples LXVII. 1. Apply Cauchy’s and d’Alembert’s tests (as specialised in 4 above) to the series nkrn, where k is a positive rational number.

[Here vn+1/vn = {(n + 1)/n}kr r, so that d’Alembert’s test shows at once that the series is convergent if r < 1 and divergent if r > 1. The test fails if r = 1: but the series is then obviously divergent. Since limn1/n = 1 (Ex. XXVII. 11), Cauchy’s test leads at once to the same conclusions.]

2. Consider the series (Ank + Bnk1 + + K)rn. [We may suppose A positive. If the coefficient of rn is denoted by P(n), then P(n)/nk A and, by D of §167, the series behaves like nkrn.]

3. Consider

Ank + Bnk1 + + K αnl + βnl1 + + κ rn(A > 0,α > 0).

[The series behaves like nklrn. The case in which r = 1, k < l requires further consideration.]

4. We have seen (Ch. IV, Misc. Ex. 17) that the series

1 n(n + 1), 1 n(n + 1)(n + p)
are convergent. Show that Cauchy’s and d’Alembert’s tests both fail when applied to them. [For limun1/n = lim(un+1/un) = 1.]

5. Show that the series  np, where p is an integer not less than 2, is convergent. [Since lim{n(n + 1)(n + p 1)}/np = 1, this follows from the convergence of the series considered in Ex. 4. It has already been shown in §77, (7) that the series is divergent if p = 1, and it is obviously divergent if p 0.]

6. Show that the series

Ank + Bnk1 + + K αnl + βnl1 + + κ
is convergent if l > k + 1 and divergent if l k + 1.

7. If mn is a positive integer, and mn+1 > mn, then the series rmn is convergent if r < 1 and divergent if r 1. For example the series 1 + r + r4 + r9 + is convergent if r < 1 and divergent if r 1.

8. Sum the series 1 + 2r + 2r4 + to 24 places of decimals when r = .1 and to 2 places when r = .9. [If r = .1, then the first 5 terms give the sum 1.2002000020000002, and the error is

2r25 + 2r36 + < 2r25 + 2r36 + 2r47 + = 2r25/(1 r11) < 3/1025.
If r = .9, then the first 8 terms give the sum 5.458, and the error is less than 2r64/(1 r17) < .003.]

9. If 0 < a < b < 1, then the series a + b + a2 + b2 + a3 + is convergent. Show that Cauchy’s test may be applied to this series, but that d’Alembert’s test fails. [For

v2n+1/v2n = (b/a)n+1 ,v 2n+2/v2n+1 = b(a/b)n+2 0.]

10. The series 1 + r + r2 2! + r3 3! + and 1 + r + r2 22 + r3 33 + are convergent for all positive values of r.

11. If un is convergent then so are un2 and un/(1 + un).

12. If un2 is convergent then so is un/n. [For 2un/n un2 + (1/n2) and (1/n2) is convergent.]

13. Show that

1 + 1 32 + 1 52 + = 3 4 1 + 1 22 + 1 32 +
and
1 + 1 22 + 1 32 + 1 52 + 1 62 + 1 72 + 1 92 + = 15 16 1 + 1 22 + 1 32 + .

[To prove the first result we note that

1 + 1 22 + 1 32 + = 1 + 1 22 + 1 32 + 1 42 + = 1 + 1 32 + 1 52 + + 1 22 1 + 1 22 + 1 32 + ,

by theorems (8) and (6) of §77.]

14. Prove by a reductio ad absurdum that (1/n) is divergent. [If the series were convergent we should have, by the argument used in Ex. 13,

1 + 1 2 + 1 3 + = (1 + 1 3 + 1 5 + ) + 1 2(1 + 1 2 + 1 3 + ),
or
1 2 + 1 4 + 1 6 + = 1 + 1 3 + 1 5 +
which is obviously absurd, since every term of the first series is less than the corresponding term of the second.]

169. Before proceeding further in the investigation of tests of convergence and divergence, we shall prove an important general theorem concerning series of positive terms.

Dirichlet’s Theorem.91 The sum of a series of positive terms is the same in whatever order the terms are taken.

This theorem asserts that if we have a convergent series of positive terms, u0 + u1 + u2 + say, and form any other series

v0 + v1 + v2 +
out of the same terms, by taking them in any new order, then the second series is convergent and has the same sum as the first. Of course no terms must be omitted: every u must come somewhere among the vs, and vice versa.

The proof is extremely simple. Let s be the sum of the series of us. Then the sum of any number of terms, selected from the us, is not greater than s. But every v is a u, and therefore the sum of any number of terms selected from the vs is not greater than s. Hence vn is convergent, and its sum t is not greater than s. But we can show in exactly the same way that s t. Thus s = t.

170. Multiplication of Series of Positive Terms. An immediate corollary from Dirichlet’s Theorem is the following theorem: if u0 + u1 + u2 + and v0 + v1 + v2 + are two convergent series of positive terms, and s and t are their respective sums, then the series

u0v0 + (u1v0 + u0v1) + (u2v0 + u1v1 + u0v2) +
is convergent and has the sum st.

Arrange all the possible products of pairs umvn in the form of a doubly infinite array

pict
We can rearrange these terms in the form of a simply infinite series in a variety of ways. Among these are the following.

(1) We begin with the single term u0v0 for which m + n = 0; then we take the two terms u1v0u0v1 for which m + n = 1; then the three terms u2v0u1v1u0v2 for which m + n = 2; and so on. We thus obtain the series

u0v0 + (u1v0 + u0v1) + (u2v0 + u1v1 + u0v2) +
of the theorem.

(2) We begin with the single term u0v0 for which both suffixes are zero; then we take the terms u1v0u1v1u0v1 which involve a suffix 1 but no higher suffix; then the terms u2v0, u2v1, u2v2, u1v2u0v2 which involve a suffix 2 but no higher suffix; and so on. The sums of these groups of terms are respectively equal to

u0v0,(u0 + u1)(v0 + v1) u0v0, (u0 + u1 + u2)(v0 + v1 + v2) (u0 + u1)(v0 + v1),

and the sum of the first n + 1 groups is

(u0 + u1 + + un)(v0 + v1 + + vn),
and tends to st as n . When the sum of the series is formed in this manner the sum of the first one, two, three, … groups comprises all the terms in the first, second, third, … rectangles indicated in the diagram above.

The sum of the series formed in the second manner is st. But the first series is (when the brackets are removed) a rearrangement of the second; and therefore, by Dirichlet’s Theorem, it converges to the sum st. Thus the theorem is proved.

Examples LXVIII. 1. Verify that if r < 1 then

1 + r2 + r + r4 + r6 + r3 + = 1 + r + r3 + r2 + r5 + r7 + = 1/(1 r).

2.92If either of the series u0 + u1 + , v0 + v1 + is divergent, then so is the series u0v0 + (u1v0 + u0v1) + (u2v0 + u1v1 + u0v2) + , except in the trivial case in which every term of one series is zero.

3. If the series u0 + u1 + , v0 + v1 + , w0 + w1 + converge to sums rst, then the series λk, where λk = umvnwp, the summation being extended to all sets of values of mnp such that m + n + p = k, converges to the sum rst.

4. If un and  vn converge to sums s and t, then the series  wn, where wn = ulvm, the summation extending to all pairs lm for which lm = n, converges to the sum st.

171. Further tests for convergence and divergence. The examples on pp. 10491054 suffice to show that there are simple and interesting types of series of positive terms which cannot be dealt with by the general tests of § 168. In fact, if we consider the simplest type of series, in which un+1/un tends to a limit as n , the tests of § 168 generally fail when this limit is 1. Thus in Ex. LXVII. 5 these tests failed, and we had to fall back upon a special device, which was in essence that of using the series of Ex. LXVII. 4 as our comparison series, instead of the geometric series.

The fact is that the geometric series, by comparison with which the tests of §168 were obtained, is not only convergent but very rapidly convergent, far more rapidly than is necessary in order to ensure convergence. The tests derived from comparison with it are therefore naturally very crude, and much more delicate tests are often wanted.

We proved in Ex. XXVII. 7 that nkrn 0 as n , provided r < 1, whatever value k may have; and in Ex. LXVII. 1 we proved more than this, viz. that the series nkrn is convergent. It follows that the sequence rr2, r3, …, rn, …, where r < 1, diminishes more rapidly than the sequence 1k2k, 3k, …, nk, …. This seems at first paradoxical if r is not much less than unity, and k is large. Thus of the two sequences

2 3,4 9, 8 27,;1, 1 4096, 1 531,441,
whose general terms are (2 3)n and n12, the second seems at first sight to decrease far more rapidly. But this is far from being the case; if only we go far enough into the sequences we shall find the terms of the first sequence very much the smaller. For example,
(2/3)4 = 16/81 < 1/5,(2/3)12 < (1/5)3 < (1/10)2,(2/3)1000 < (1/10)166,
while
100012 = 1036;
so that the 1000th term of the first sequence is less than the 10130th part of the corresponding term of the second sequence. Thus the series (2/3)n is far more rapidly convergent than the series n12, and even this series is very much more rapidly convergent than  n2.93

172. We shall proceed to establish two further tests for the convergence or divergence of series of positive terms, Maclaurin’s (or Cauchy’s) Integral Test and Cauchy’s Condensation Test, which, though very far from being completely general, are sufficiently general for our needs in this chapter.

In applying either of these tests we make a further assumption as to the nature of the function un, about which we have so far assumed only that it is positive. We assume that un decreases steadily with n: i.e. that un+1 un for all values of n, or at any rate all sufficiently large values.

This condition is satisfied in all the most important cases. From one point of view it may be regarded as no restriction at all, so long as we are dealing with series of positive terms: for in virtue of Dirichlet’s theorem above we may rearrange the terms without affecting the question of convergence or divergence; and there is nothing to prevent us rearranging the terms in descending order of magnitude, and applying our tests to the series of decreasing terms thus obtained.

But before we proceed to the statement of these two tests, we shall state and prove a simple and important theorem, which we shall call Abel’s Theorem.94 This is a one-sided theorem in that it gives a sufficient test for divergence only and not for convergence, but it is essentially of a more elementary character than the two theorems mentioned above.

173. Abel’s (or Pringsheim’s) Theorem. If un is a convergent series of positive and decreasing terms, then limnun = 0.

Suppose that nun does not tend to zero. Then it is possible to find a positive number δ such that nun δ for an infinity of values of n. Let n1 be the first such value of n; n2 the next such value of n which is more than twice as large as n1; n3 the next such value of n which is more than twice as large as n2; and so on. Then we have a sequence of numbers n1n2, n3, … such that n2 > 2n1, n3 > 2n2, … and so n2 n1 > 1 2n2, n3 n2 > 1 2n3, …; and also n1un1 δ, n2un2 δ, …. But, since un decreases as n increases, we have

u0 + u1 + + un11 n1un1 δ, un1 + + un21 (n2 n1)un2 > 1 2n2un2 1 2δ, un2 + + un31 (n3 n2)un3 > 1 2n3un3 1 2δ,

and so on. Thus we can bracket the terms of the series  un so as to obtain a new series whose terms are severally greater than those of the divergent series

δ + 1 2δ + 1 2δ + ;
and therefore un is divergent.

Examples LXIX. 1. Use Abel’s theorem to show that (1/n) and {1/(an + b)} are divergent. [Here nun 1 or nun 1/a.]

2. Show that Abel’s theorem is not true if we omit the condition that un decreases as n increases. [The series

1 + 1 22 + 1 32 + 1 4 + 1 52 + 1 62 + 1 72 + 1 82 + 1 9 + 1 102 + ,
in which un = 1/n or 1/n2, according as n is or is not a perfect square, is convergent, since it may be rearranged in the form
1 22 + 1 32 + 1 52 + 1 62 + 1 72 + 1 82 + 1 102 + + 1 + 1 4 + 1 9 + ,
and each of these series is convergent. But, since nun = 1 whenever n is a perfect square, it is clearly not true that nun 0.]

3. The converse of Abel’s theorem is not true, i.e. it is not true that, if un decreases with n and limnun = 0, then un is convergent.

[Take the series (1/n) and multiply the first term by 1, the second by 1 2, the next two by 1 3, the next four by 1 4, the next eight by 1 5, and so on. On grouping in brackets the terms of the new series thus formed we obtain

1 + 1 2 1 2 + 1 3 1 3 + 1 4 + 1 4 1 5 + 1 6 + 1 7 + 1 8 + ;
and this series is divergent, since its terms are greater than those of
1 + 1 2 1 2 + 1 3 1 2 + 1 4 1 2 + ,
which is divergent. But it is easy to see that the terms of the series
1 + 1 2 1 2 + 1 3 1 3 + 1 3 1 4 + 1 4 1 5 + 1 4 1 6 +
satisfy the condition that nun 0. In fact nun = 1/ν if 2ν2 < n 2ν1, and ν as n .]

174. Maclaurin’s (or Cauchy’s) Integral Test.95 If un decreases steadily as n increases, we can write un = φ(n) and suppose that φ(n) is the value assumed, when x = n, by a continuous and steadily decreasing function φ(x) of the continuous variable x. Then, If ν is any positive integer, we have

φ(ν 1) φ(x) φ(ν)
when ν 1 x ν. Let
vν = φ(ν 1) ν1νφ(x)dx =ν1ν{φ(ν 1) φ(x)}dx,
so that
0 vν φ(ν 1) φ(ν).
Then vν is a series of positive terms, and
v2 + v3 + + vn φ(1) φ(n) φ(1).
Hence vν is convergent, and so v2 + v3 + + vn or
1n1φ(ν) 1nφ(x)dx
tends to a positive limit as n .

Let us write

Φ(ξ) =1ξφ(x)dx,
so that Φ(ξ) is a continuous and steadily increasing function of ξ. Then
u1 + u2 + + un1 Φ(n)
tends to a positive limit, not greater than φ(1), as n . Hence uν is convergent or divergent according as Φ(n) tends to a limit or to infinity as n , and therefore, since Φ(n) increases steadily, according as Φ(ξ) tends to a limit or to infinity as ξ . Hence if φ(x) is a function of x which is positive and continuous for all values of x greater than unity, and decreases steadily as x increases, then the series
φ(1) + φ(2) +
does or does not converge according as
Φ(ξ) =1ξφ(x)dx
does or does not tend to a limit l as ξ ; and, in the first case, the sum of the series is not greater than φ(1) + l.

The sum must in fact be less than φ(1) + l. For it follows from (6) of §160, and Ch. VII, Misc. Ex. 41, that vν < φ(ν 1) φ(ν), unless φ(x) = φ(ν) throughout the interval [ν 1,ν]; and this cannot be true for all values of ν.

Examples LXX. 1. Prove that

1 1 n2 + 1 < 1 2 + 1 4π.

2. Prove that

1 2π < 1 a a2+n2 < 1 2π.
(Math. Trip. 1909.)

3. Prove that if m > 0 then

1 m2 + 1 (m + 1)2 + 1 (m + 2)2 + < m + 1 m .

175. The series ns. By far the most important application of the Integral Test is to the series

1s + 2s + 3s + + ns + ,
where s is any rational number. We have seen already (§ 77 and Exs. LXVII. 14, LXIX. 1) that the series is divergent when s = 1.

If s 0 then it is obvious that the series is divergent. If s > 0 then un decreases as n increases, and we can apply the test. Here

Φ(ξ) =1ξdx xs = ξ1s 1 1 s ,
unless s = 1. If s > 1 then ξ1s 0 as ξ , and
Φ(ξ) 1 (s 1) = l,
say. And if s < 1 then ξ1s as ξ , and so Φ(ξ) . Thus the series ns is convergent if s > 1, divergent if s 1, and in the first case its sum is less than s/(s 1).

So far as divergence for s < 1 is concerned, this result might have been derived at once from comparison with  (1/n), which we already know to be divergent.

It is however interesting to see how the Integral Test may be applied to the series  (1/n), when the preceding analysis fails. In this case

Φ(ξ) =1ξdx x ,
and it is easy to see that Φ(ξ) as ξ . For if ξ > 2n then
Φ(ξ) >12n dx x =12dx x +24dx x + +2n12n dx x .
But by putting x = 2ru we obtain
2r2r+1 dx x =12du u ,
and so Φ(ξ) > n12du u , which shows that Φ(ξ) as ξ .

Examples LXXI. 1. Prove by an argument similar to that used above, and without integration, that Φ(ξ) =1ξdx xs , where s < 1, tends to infinity with ξ.

2. The series n2, n3/2, n11/10 are convergent, and their sums are not greater than 2311 respectively. The series n1/2, n10/11 are divergent.

3. The series ns/(nt + a), where a > 0, is convergent or divergent according as t > 1 + s or t 1 + s. [Compare with  nst.]

4. Discuss the convergence or divergence of the series

(a1ns1 + a2ns2 + + aknsk )/(b1nt1 + b2nt2 + + blntl ),
where all the letters denote positive numbers and the s’s and t’s are rational and arranged in descending order of magnitude.

5. Prove that

2n 2 < 1 1 + 1 2 + + 1 n < 2n 1, 1 2π < 1 21 + 1 32 + 1 43 + < 1 2(π + 1).

(Math. Trip. 1911.)

6. If φ(n) l > 1 then the series nφ(n) is convergent. If φ(n) l < 1 then it is divergent.

176. Cauchy’s Condensation Test. The second of the two tests mentioned in § 172 is as follows: if un = φ(n) is a decreasing function of n, then the series φ(n) is convergent or divergent according as 2nφ(2n) is convergent or divergent.

We can prove this by an argument which we have used already (§ 77) in the special case of the series (1/n). In the first place

φ(3) + φ(4) 2φ(4), φ(5) + φ(6) + + φ(8) 4φ(8), .................................. φ(2n + 1) + φ(2n + 2) + + φ(2n+1) 2nφ(2n+1).

If 2nφ(2n) diverges then so do 2n+1φ(2n+1) and 2nφ(2n+1), and then the inequalities just obtained show that φ(n) diverges.

On the other hand

φ(2) + φ(3) 2φ(2),φ(4) + φ(5) + + φ(7) 4φ(4),
and so on. And from this set of inequalities it follows that if 2nφ(2n) converges then so does φ(n). Thus the theorem is established.

For our present purposes the field of application of this test is practically the same as that of the Integral Test. It enables us to discuss the series ns with equal ease. For ns will converge or diverge according as 2n2ns converges or diverges, i.e. according as s > 1 or s 1.

Examples LXXII. 1. Show that if a is any positive integer greater than 1 then φ(n) is convergent or divergent according as anφ(an) is convergent or divergent. [Use the same arguments as above, taking groups of aa2, a3, … terms.]

2. If 2nφ(2n) converges then it is obvious that lim2nφ(2n) = 0. Hence deduce Abel’s Theorem of §173.

177. Infinite Integrals. The Integral Test of § 174 shows that, if φ(x) is a positive and decreasing function of x, then the series φ(n) is convergent or divergent according as the integral function Φ(x) does or does not tend to a limit as x . Let us suppose that it does tend to a limit, and that

lim x1xφ(t)dt = l.
Then we shall say that the integral
1φ(t)dt
is convergent, and has the value l
; and we shall call the integral an infinite integral.

So far we have supposed φ(t) positive and decreasing. But it is natural to extend our definition to other cases. Nor is there any special point in supposing the lower limit to be unity. We are accordingly led to formulate the following definition:

If φ(t) is a function of t continuous when t a, and

lim xaxφ(t)dt = l,
then we shall say that the infinite integral

aφ(t)dt (1)

is convergent and has the value l.

The ordinary integral between limits a and A, as defined in Ch. VII, we shall sometimes call in contrast a finite integral.

On the other hand, when

axφ(t)dt ,
we shall say that the integral diverges to , and we can give a similar definition of divergence to  . Finally, when none of these alternatives occur, we shall say that the integral oscillates, finitely or infinitely, as x .

These definitions suggest the following remarks.

(i) If we write

axφ(t)dt = Φ(x),
then the integral converges, diverges, or oscillates according as Φ(x) tends to a limit, tends to  (or to  ), or oscillates, as x . If Φ(x) tends to a limit, which we may denote by Φ(), then the value of the integral is Φ(). More generally, if Φ(x) is any integral function of φ(x), then the value of the integral is Φ() Φ(a).

(ii) In the special case in which φ(t) is always positive it is clear that Φ(x) is an increasing function of x. Hence the only alternatives are convergence and divergence to .

(iii) The integral (1) of course depends on a, but is quite independent of t, and is in no way altered by the substitution of any other letter for t (cf. §157).

(iv) Of course the reader will not be puzzled by the use of the term infinite integral to denote something which has a definite value such as 2 or 1 2π. The distinction between an infinite integral and a finite integral is similar to that between an infinite series and a finite series: no one supposes that an infinite series is necessarily divergent.

(v) The integral axφ(t)dt was defined in §§156 and 157 as a simple limit, i.e. the limit of a certain finite sum. The infinite integral is therefore the limit of a limit, or what is known as a repeated limit. The notion of the infinite integral is in fact essentially more complex than that of the finite integral, of which it is a development.

(vi) The Integral Test of §174 may now be stated in the form: if φ(x) is positive and steadily decreases as x increases, then the infinite series φ(n) and the infinite integral 1φ(x)dx converge or diverge together.

(vii) The reader will find no difficulty in formulating and proving theorems for infinite integrals analogous to those stated in (1)–(6) of §77. Thus the result analogous to (2) is that if aφ(x)dx is convergent, and b > a, then bφ(x)dx is convergent and

aφ(x)dx =abφ(x)dx +bφ(x)dx.

178. The case in which φ(x) is positive. It is natural to consider what are the general theorems, concerning the convergence or divergence of the infinite integral (1) of § 177, analogous to theorems A–D of § 167. That A is true of integrals as well as of series we have already seen in § 177, (ii). Corresponding to B we have the theorem that the necessary and sufficient condition for the convergence of the integral (1) is that it should be possible to find a constant K such that

axφ(t)dt < K
for all values of x greater than a.

Similarly, corresponding to C, we have the theorem: if aφ(x)dx is convergent, and ψ(x) Kφ(x) for all values of x greater than a, then aψ(x)dx is convergent and

aψ(x)dx Kaφ(x)dx.
We leave it to the reader to formulate the corresponding test for divergence.

We may observe that d’Alembert’s test (§ 168), depending as it does on the notion of successive terms, has no analogue for integrals; and that the analogue of Cauchy’s test is not of much importance, and in any case could only be formulated when we have investigated in greater detail the theory of the function φ(x) = rx, as we shall do in Ch. IX. The most important special tests are obtained by comparison with the integral

adx xs (a > 0),
whose convergence or divergence we have investigated in § 175, and are as follows: if φ(x) < Kxs, where s > 1, when x a, then aφ(x)dx is convergent; and if φ(x) > Kxs, where s 1, when x a, then the integral is divergent; and in particular, if lim xsφ(x) = l, where l > 0, then the integral is convergent or divergent according as s > 1 or s 1.

There is one fundamental property of a convergent infinite series in regard to which the analogy between infinite series and infinite integrals breaks down. If φ(n) is convergent then φ(n) 0; but it is not always true, even when φ(x) is always positive, that if aφ(x)dx is convergent then φ(x) 0.

Consider for example the function φ(x) whose graph is indicated by the thick line in the figure. Here the height of the peaks corresponding to the points x = 1, 23, … is in each case unity, and the breadth of the peak corresponding


pict

Fig. 50.

to x = n is 2/(n + 1)2. The area of the peak is 1/(n + 1)2, and it is evident that, for any value of ξ,

0ξφ(x)dx < 0 1 (n + 1)2,
so that 0φ(x)dx is convergent; but it is not true that φ(x) 0.

Examples LXXIII. 1. The integral

aαxr + βxr1 + + λ Axs + Bxs1 + + Ldx,
where α and A are positive and a is greater than the greatest root of the denominator, is convergent if s > r + 1 and otherwise divergent.

2. Which of the integrals adx x, a dx x4/3,

a dx c2 + x2,a xdx c2 + x2,a x2dx c2 + x2,a x2dx α + 2βx2 + γx4
are convergent? In the first two integrals it is supposed that a > 0, and in the last that a is greater than the greatest root (if any) of the denominator.

3. The integrals

aξ cosxdx,aξ sinxdx,aξ cos(αx + β)dx
oscillate finitely as ξ .

4. The integrals

aξxcosxdx,aξx2 sinxdxaξxn cos(αx + β)dx,
where n is any positive integer, oscillate infinitely as ξ .

5. Integrals to  . If ξaφ(x)dx tends to a limit l as ξ , then we say that aφ(x)dx is convergent and equal to l. Such integrals possess properties in every respect analogous to those of the integrals discussed in the preceding sections: the reader will find no difficulty in formulating them.

6. Integrals from  to  + . If the integrals

aφ(x)dx,aφ(x)dx
are both convergent, and have the values kl respectively, then we say that
φ(x)dx
is convergent and has the value k + l.

7. Prove that

0 dx 1 + x2 =0 dx 1 + x2 = 1 2 dx 1+x2 = 1 2π.

8. Prove generally that

φ(x2)dx = 20φ(x2)dx,
provided that the integral 0φ(x2)dx is convergent.

9. Prove that if 0xφ(x2)dx is convergent then xφ(x2)dx = 0.

10. Analogue of Abel’s Theorem of §173. If φ(x) is positive and steadily decreases, and aφ(x)dx is convergent, then xφ(x) 0. Prove this (a) by means of Abel’s Theorem and the Integral Test and (b) directly, by arguments analogous to those of §173.

11. If a = x0 < x1 < x2 < and xn , and un =xnxn+1 φ(x)dx, then the convergence of aφ(x)dx involves that of un. If φ(x) is always positive the converse statement is also true. [That the converse is not true in general is shown by the example in which φ(x) = cosx, xn = nπ.]

179. Application to infinite integrals of the rules for substitution and integration by parts. The rules for the transformation of a definite integral which were discussed in § 161 may be extended so as to apply to infinite integrals.

(1) Transformation by substitution. Suppose that

aφ(x)dx (1)

is convergent. Further suppose that, for any value of ξ greater than a, we have, as in § 161,

aξφ(x)dx =bτφ{f(t)}f(t)dt, (2)

where a = f(b), ξ = f(τ). Finally suppose that the functional relation x = f(t) is such that x as t . Then, making τ and so ξ tend to  in (2), we see that the integral

bφ{f(t)}f(t)dt (3)

is convergent and equal to the integral (1).

On the other hand it may happen that ξ as τ or as τ c. In the first case we obtain

aφ(x)dx = lim τbτφ{f(t)}f(t)dt = lim ττbφ{f(t)}f(t)dt = bφ{f(t)}f(t)dt.

In the second case we obtain

aφ(x)dx = lim τcbτφ{f(t)}f(t)dt. (4)

We shall return to this equation in § 181.

There are of course corresponding results for the integrals

aφ(x)dx,φ(x)dx,
which it is not worth while to set out in detail: the reader will be able to formulate them for himself.

Examples LXXIV. 1. Show, by means of the substitution x = tα, that if s > 1 and α > 0 then

1xsdx = α1tα(1s)1dt;
and verify the result by calculating the value of each integral directly.

2. If aφ(x)dx is convergent then it is equal to one or other of

α(aβ)/αφ(αt + β)dt, α(aβ)/αφ(αt + β)dt,
according as α is positive or negative.

3. If φ(x) is a positive and steadily decreasing function of x, and α and β are any positive numbers, then the convergence of the series φ(n) implies and is implied by that of the series φ(αn + β).

[It follows at once, on making the substitution x = αt + β, that the integrals

aφ(x)dx,(aβ)/αφ(αt + β)dt
converge or diverge together. Now use the Integral Test.]

4. Show that

1 dx (1 + x)x = 1 2π.

[Put x = t2.]

5. Show that

0 x (1 + x)2dx = 1 2π.

[Put x = t2 and integrate by parts.]

6. If φ(x) h as x , and φ(x) k as x , then

{φ(x a) φ(x b)}dx = (a b)(h k).

[For

ξξ{φ(x a) φ(x b)}dx =ξξφ(x a)dx ξξφ(x b)dx =ξaξaφ(t)dt ξbξbφ(t)dt =ξaξbφ(t)dt ξaξbφ(t)dt.

The first of these two integrals may be expressed in the form

(a b)k +ξaξbρdt,
where ρ 0 as ξ, and the modulus of the last integral is less than or equal to a bκ, where κ is the greatest value of ρ throughout the interval [ξa,ξb]. Hence
ξaξbφ(t)dt (a b)k.
The second integral may be discussed similarly.]

(2) Integration by parts. The formula for integration by parts (§ 161) is

aξf(x)φ(x)dx = f(ξ)φ(ξ) f(a)φ(a) aξf(x)φ(x)dx.

Suppose now that ξ . Then if any two of the three terms in the above equation which involve ξ tend to limits, so does the third, and we obtain the result

af(x)φ(x)dx = lim ξf(ξ)φ(ξ) f(a)φ(a) af(x)φ(x)dx.
There are of course similar results for integrals to  , or from to .

Examples LXXV. 1. Show that

0 x (1 + x)3dx = 1 20 dx (1+x)2 = 1 2.

2. 0 x2 (1 + x)4dx = 2 30 x (1+x)3dx = 1 3.

3. If m and n are positive integers, and

Im,n =0 xmdx (1 + x)m+n,
then
Im,n = {m/(m + n 1)}Im1,n.
Hence prove that Im,n = m!(n 2)!/(m + n 1)!.

4. Show similarly that if

Im,n =0 x2m+1dx (1 + x2)m+n
then
Im,n = {m/(m + n 1)}Im1,n,2Im,n = m!(n 2)!/(m + n 1)!.
Verify the result by applying the substitution x = t2 to the result of Ex. 3.

180. Other types of infinite integrals. It was assumed, in the definition of the ordinary or finite integral given in Ch. VII, that (1) the range of integration is finite and (2) the subject of integration is continuous.

It is possible, however, to extend the notion of the ‘definite integral’ so as to apply to many cases in which these conditions are not satisfied. The ‘infinite’ integrals which we have discussed in the preceding sections, for example, differ from those of Ch. VII in that the range of integration is infinite. We shall now suppose that it is the second of the conditions (1), (2) that is not satisfied. It is natural to try to frame definitions applicable to some such cases at any rate. There is only one such case which we shall consider here. We shall suppose that φ(x) is continuous throughout the range of integration [a,A] except for a finite number of values of x, say x = ξ1, ξ2, …, and that φ(x) or φ(x) as x tends to any of these exceptional values from either side.

It is evident that we need only consider the case in which [a,A] contains one such point ξ. When there is more than one such point we can divide up [a,A] into a finite number of sub-intervals each of which contains only one; and, if the value of the integral over each of these sub-intervals has been defined, we can then define the integral over the whole interval as being the sum of the integrals over each sub-interval. Further, we can suppose that the one point ξ in [a,A] comes at one or other of the limits aA. For, if it comes between a and A, we can then define aAφ(x)dx as

aξφ(x)dx +ξAφ(x)dx,
assuming each of these integrals to have been satisfactorily defined. We shall suppose, then, that ξ = a; it is evident that the definitions to which we are led will apply, with trifling changes, to the case in which ξ = A.

Let us then suppose φ(x) to be continuous throughout [a,A] except for x = a, while φ(x) as x a through values greater than a. A typical example of such a function is given by

φ(x) = (x a)s,
where s > 0; or, in particular, if a = 0, by φ(x) = xs. Let us therefore consider how we can define

0Adx xs , (1)

when s > 0.

The integral 1/Ays2dy is convergent if s < 1 (§ 175) and means lim η1/Aηys2dy. But if we make the substitution y = 1/x, we obtain

1/Aηys2dy =1/ηAxsdx.
Thus lim η1/ηAxsdx, or, what is the same thing,
lim ε+0εAxsdx,
exists provided that s < 1; and it is natural to define the value of the integral (1) as being equal to this limit. Similar considerations lead us to define aA(x a)sdx by the equation
aA(x a)sdx = lim ε+0a+εA(x a)sdx.

We are thus led to the following general definition: if the integral

a+εAφ(x)dx
tends to a limit l as ε +0, we shall say that the integral
aAφ(x)dx
is convergent and has the value l.

Similarly, when φ(x) as x tends to the upper limit A, we define aAφ(x)dx as being

lim ε+0aAεφ(x)dx :
and then, as we explained above, we can extend our definitions to cover the case in which the interval [a,A] contains any finite number of infinities of φ(x).

An integral in which the subject of integration tends to  or to  as x tends to some value or values included in the range of integration will be called an infinite integral of the second kind: the first kind of infinite integrals being the class discussed in §§ 177 et seq. Nearly all the remarks (i)–(vii) made at the end of § 177 apply to infinite integrals of the second kind as well as to those of the first.

181. We may now write the equation (4) of §179 in the form

aφ(x)dx =bcφ{f(t)}f(t)dt. (1)

The integral on the right-hand side is defined as the limit, as τ c, of the corresponding integral over the range [b,τ], i.e. as an infinite integral of the second kind. And when φ{f(t)}f(t) has an infinity at t = c the integral is essentially an infinite integral. Suppose for example, that φ(x) = (1 + x)m, where 1 < m < 2, and a = 0, and that f(t) = t/(1 t). Then b = 0, c = 1, and (1) becomes

0 dx (1 + x)m =01(1 t)m2dt; (2)

and the integral on the right-hand side is an infinite integral of the second kind.

On the other hand it may happen that φ{f(t)}f(t) is continuous for t = c. In this case

bcφ{f(t)}f(t)dt
is a finite integral, and
limτcbτφ{f(t)}f(t)dt =bcφ{f(t)}f(t)dt,
in virtue of the corollary to Theorem (10) of §160. In this case the substitution x = f(t) transforms an infinite into a finite integral. This case arises if m 2 in the example considered a moment ago.

Examples LXXVI. 1. If φ(x) is continuous except for x = a, while φ(x) as x a, then the necessary and sufficient condition that aAφ(x)dx should be convergent is that we can find a constant K such that

a+εAφ(x)dx < K
for all values of ε, however small (cf. §178).

It is clear that we can choose a number A between a and A, such that φ(x) is positive throughout [a,A]. If φ(x) is positive throughout the whole interval [a,A] then we can of course identify A and A. Now

aεAφ(x)dx =aεAφ(x)dx +AAφ(x)dx.
The first integral on the right-hand side of the above equation increases as ε decreases, and therefore tends to a limit or to ; and the truth of the result stated becomes evident.

If the condition is not satisfied then aεAφ(x)dx . We shall then say that the integral aAφ(x)dx diverges to . It is clear that, if φ(x) as x a + 0, then convergence and divergence to  are the only alternatives for the integral. We may discuss similarly the case in which φ(x) .

2. Prove that

aA(x a)sdx = (A a)1s 1 s
if s < 1, while the integral is divergent if s 1.

3. If φ(x) as x a + 0 and φ(x) < K(x a)s, where s < 1, then aAφ(x)dx is convergent; and if φ(x) > K(x a)s, where s 1, then the integral is divergent. [This is merely a particular case of a general comparison theorem analogous to that stated in §178.]

4. Are the integrals

aA dx (x a)(A x),aA dx (A x)x a3,aA dx (A x)A x3, aA dx x2 a2,aA dx A3 x33,aA dx x2 a2,aA dx A3 x3

convergent or divergent?

5. The integrals

11dx x3,a1a+1 dx x a3
are convergent, and the value of each is zero.

6. The integral

0π dx sinx
is convergent. [The subject of integration tends to as x tends to either limit.]

7. The integral

0π dx (sinx)s
is convergent if and only if s < 1.

8. The integral

01 2π xs (sinx)tdx
is convergent if t < s + 1.

9. Show that

0h sinx xp dx,
where h > 0, is convergent if p < 2. Show also that, if 0 < p < 2, the integrals
0π sinx xp dx,π2π sinx xp dx,2π3π sinx xp dx,
alternate in sign and steadily decrease in absolute value. [Transform the integral whose limits are kπ and (k + 1)π by the substitution x = kπ + y.]

10. Show that

0h sinx xp dx,
where 0 < p < 2, attains its greatest value when h = π.
(Math. Trip. 1911.)

11. The integral

01 2π(cosx)l(sinx)mdx
is convergent if and only if l > 1, m > 1.

12. Such an integral as

0xs1dx 1 + x ,
where s < 1, does not fall directly under any of our previous definitions. For the range of integration is infinite and the subject of integration tends to as x +0. It is natural to define this integral as being equal to the sum
01xs1dx 1 + x +1xs1dx 1 + x ,
provided that these two integrals are both convergent.

The first integral is a convergent infinite integral of the second kind if 0 < s < 1. The second is a convergent infinite integral of the first kind if s < 1. It should be noted that when s > 1 the first integral is an ordinary finite integral; but then the second is divergent. Thus the integral from 0 to is convergent if and only if 0 < s < 1.

13. Prove that

0 xs1 1 + xtdx
is convergent if and only if 0 < s < t.

14. The integral

0xs1 xt1 1 x dx
is convergent if and only if 0 < s < 1, 0 < t < 1. [It should be noticed that the subject of integration is undefined when x = 1; but (xs1 xt1)/(1 x) t s as x 1 from either side; so that the subject of integration becomes a continuous function of x if we assign to it the value t s when x = 1.

It often happens that the subject of integration has a discontinuity which is due simply to a failure in its definition at a particular point in the range of integration, and can be removed by attaching a particular value to it at that point. In this case it is usual to suppose the definition of the subject of integration completed in this way. Thus the integrals

01 2π sinmx x dx,01 2π sinmx sinx dx
are ordinary finite integrals, if the subjects of integration are regarded as having the value m when x = 0.]

15. Substitution and integration by parts. The formulae for transformation by substitution and integration by parts may of course be extended to infinite integrals of the second as well as of the first kind. The reader should formulate the general theorems for himself, on the lines of §179.

16. Prove by integration by parts that if s > 0, t > 1, then

01xs1(1 x)t1dx = t 1 s 01xs(1 x)t2dx.

17. If s > 0 then

01xs1dx 1 + x =1tsdt 1 + t.

[Put x = 1/t.]

18. If 0 < s < 1 then

01xs1 + xs 1 + x dx =0tsdt 1 + t =0ts1dt 1 + t .

19. If a + b > 0 then

b dx (x + a)x b = π a + b.
(Math. Trip. 1909.)

20. Show, by means of the substitution x = t/(1 t), that if l and m are both positive then

0 xl1 (1 + x)l+mdx =01tl1(1 t)m1dt.

21. Show, by means of the substitution x = pt/(p + 1 t), that if lm, and p are all positive then

01xl1(1 x)m1 dx (x + p)l+m = 1 (1 + p)lpm01tl1(1 t)m1dt.

22. Prove that

ab dx (x a)(b x) = π andab xdx (x a)(b x) = 1 2π(a + b),
(i) by means of the substitution x = a + (b a)t2, (ii) by means of the substitution (b x)/(x a) = t, and (iii) by means of the substitution x = acos2t + bsin2t.

23. If s > 1 then

01 2π(sinθ)sdθ =01 xsdx 1 x2 = 1 201x1 2(s1)dx 1x = 1 201(1 x)1 2(s1) dx x.

24. Establish the formulae

01f(x)dx 1 x2 =01 2πf(sinθ)dθ, ab f(x)dx (x a)(b x) = 201 2πf(acos2θ + bsin2θ)dθ, aaf a x a + xdx = 4a01 2πf(tanθ)cosθsinθdθ.

25. Prove that

01 dx (1 + x)(2 + x)x(1 x) = π 1 2 1 6.

[Put x = sin2θ and use Ex. LXIII. 8.]

(Math. Trip. 1912.)

182. Some care has occasionally to be exercised in applying the rule for transformation by substitution. The following example affords a good illustration of this.

Let

J =17(x2 6x + 13)dx.
We find by direct integration that J = 48. Now let us apply the substitution
y = x2 6x + 13,
which gives x = 3 ±y 4. Since y = 8 when x = 1 and y = 20 when x = 7, we appear to be led to the result
J =820ydx dydy = ±1 2820 ydy y4.
The indefinite integral is
1 3(y 4)3/2 + 4(y 4)1/2,
and so we obtain the value  ±80 3 , which is certainly wrong whichever sign we choose.

The explanation is to be found in a closer consideration of the relation between x and y. The function x2 6x + 13 has a minimum for x = 3, when y = 4. As x increases from 1 to 3, y decreases from 8 to 4, and dx/dy is negative, so that

dx dy = 1 2y 4.
As x increases from 3 to 7, y increases from 4 to 20, and the other sign must be chosen. Thus
J =17ydx =84 y 2y 4dy +420 y 2y 4dy,
a formula which will be found to lead to the correct result.

Similarly, if we transform the integral 0πdx = π by the substitution x = arcsiny, we must observe that dx/dy = 1/1 y2 or dx/dy = 1/1 y2 according as 0 x < 1 2π or 1 2π < x π.

Example. Verify the results of transforming the integrals

01(4x2 x + 1 16)dx,0π cos2xdx
by the substitutions 4x2 x + 1 16 = y, x = arcsiny respectively.

183. Series of positive and negative terms. Our definitions of the sum of an infinite series, and the value of an infinite integral, whether of the first or the second kind, apply to series of terms or integrals of functions whose values may be either positive or negative. But the special tests for convergence or divergence which we have established in this chapter, and the examples by which we have illustrated them, have had reference almost entirely to the case in which all these values are positive. Of course the case in which they are all negative is not essentially different, as it can be reduced to the former by changing un into un or φ(x) into  φ(x).

In the case of a series it has always been explicitly or tacitly assumed that any conditions imposed upon un may be violated for a finite number of terms: all that is necessary is that such a condition (e.g. that all the terms are positive) should be satisfied from some definite term onwards. Similarly in the case of an infinite integral the conditions have been supposed to be satisfied for all values of x greater than some definite value, or for all values of x within some definite interval [a,a + δ] which includes the value a near which the subject of integration tends to infinity. Thus our tests apply to such a series as

n2 10 n4 ,
since n2 10 > 0 when n 4, and to such integrals as
1 3x 7 (x + 1)3dx,011 2x x dx,
since 3x 7 > 0 when x > 7 3, and 1 2x > 0 when 0 < x < 1 2.

But when the changes of sign of un persist throughout the series, i.e. when the number of both positive and negative terms is infinite, as in the series 1 1 2 + 1 3 1 4 + ; or when φ(x) continually changes sign as x , as in the integral

1 sin x xs dx,
or as x a, where a is a point of discontinuity of φ(x), as in the integral
aA sin 1 x a dx x a;
then the problem of discussing convergence or divergence becomes more difficult. For now we have to consider the possibility of oscillation as well as of convergence or divergence.

We shall not, in this volume, have to consider the more general problem for integrals. But we shall, in the ensuing chapters, have to consider certain simple examples of series containing an infinite number of both positive and negative terms.

184. Absolutely Convergent Series. Let us then consider a series un in which any term may be either positive or negative. Let

un = αn,
so that αn = un if un is positive and αn = un if un is negative. Further, let vn = un or vn = 0, according as un is positive or negative, and wn = un or wn = 0, according as un is negative or positive; or, what is the same thing, let vn or wn be equal to αn according as un is positive or negative, the other being in either case equal to zero. Then it is evident that vn and wn are always positive, and that
un = vn wn,αn = vn + wn.

If, for example, our series is 1 (1/2)2 + (1/3)2 , then un = (1)n1/n2 and αn = 1/n2, while vn = 1/n2 or vn = 0 according as n is odd or even and wn = 1/n2 or wn = 0 according as n is even or odd.

We can now distinguish two cases.

A. Suppose that the series αn is convergent. This is the case, for instance, in the example above, where αn is

1 + (1/2)2 + (1/3)2 + .
Then both vn and wn are convergent: for (Ex. XXX. 18) any series selected from the terms of a convergent series of positive terms is convergent. And hence, by theorem (6) of § 77, un or (vn wn) is convergent and equal to vn wn.

We are thus led to formulate the following definition.

DEFINITION. When αn or unis convergent, the series un is said to be absolutely convergent.

And what we have proved above amounts to this: if un is absolutely convergent then it is convergent; so are the series formed by its positive and negative terms taken separately; and the sum of the series is equal to the sum of the positive terms plus the sum of the negative terms.

The reader should carefully guard himself against supposing that the statement ‘an absolutely convergent series is convergent’ is a mere tautology. When we say that un is ‘absolutely convergent’ we do not assert directly that un is convergent: we assert the convergence of another series un, and it is by no means evident a priori that this precludes oscillation on the part of  un.

Examples LXXVII. 1. Employ the ‘general principle of convergence’ (§84) to prove the theorem that an absolutely convergent series is convergent. [Since un is convergent, we can, when any positive number ε is assigned, choose n0 so that

un1+1+ un1+2+ + un2< ε
when n2 > n1 n0. A fortiori
un1+1 + un1+2 + + un2< ε,
and therefore un is convergent.]

2. If an is a convergent series of positive terms, and bnKan, then bn is absolutely convergent.

3. If an is a convergent series of positive terms, then the series anxn is absolutely convergent when 1 x 1.

4. If an is a convergent series of positive terms, then the series an cosnθ, an sinnθ are absolutely convergent for all values of θ. [Examples are afforded by the series rn cosnθ, rn sinnθ of §88.]

5. Any series selected from the terms of an absolutely convergent series is absolutely convergent. [For the series of the moduli of its terms is a selection from the series of the moduli of the terms of the original series.]

6. Prove that if un is convergent then

un un,
and that the only case to which the sign of equality can apply is that in which every term has the same sign.

185. Extension of Dirichlet’s Theorem to absolutely convergent series. Dirichlet’s Theorem (§ 169) shows that the terms of a series of positive terms may be rearranged in any way without affecting its sum. It is now easy to see that any absolutely convergent series has the same property. For let un be so rearranged as to become un, and let αnvnwn be formed from un as αnvnwn were formed from un. Then αn is convergent, as it is a rearrangement of  αn, and so are vn, wn, which are rearrangements of vn, wn. Also, by Dirichlet’s Theorem, vn = vn and wn = wn and so

un = vn wn = vn wn = un.

186. Conditionally convergent series. B. We have now to consider the second case indicated above, viz. that in which the series of moduli αn diverges to .

DEFINITION. If un is convergent, but un divergent, the original series is said to be conditionally convergent.

In the first place we note that, if un is conditionally convergent, then the series vn, wn of § 184 must both diverge to . For they obviously cannot both converge, as this would involve the convergence of (vn + wn) or  αn. And if one of them, say wn, is convergent, and vn divergent, then

0Nu n = 0Nv n 0Nw n, (1)

and therefore tends to  with N, which is contrary to the hypothesis that un is convergent.

Hence vn, wn are both divergent. It is clear from equation (1) above that the sum of a conditionally convergent series is the limit of the difference of two functions each of which tends to  with n. It is obvious too that un no longer possesses the property of convergent series of positive terms (Ex. XXX. 18), and all absolutely convergent series (Ex. LXXVII. 5), that any selection from the terms itself forms a convergent series. And it seems more than likely that the property prescribed by Dirichlet’s Theorem will not be possessed by conditionally convergent series; at any rate the proof of § 185 fails completely, as it depended essentially on the convergence of vn and wn separately. We shall see in a moment that this conjecture is well founded, and that the theorem is not true for series such as we are now considering.

187. Tests of convergence for conditionally convergent series. It is not to be expected that we should be able to find tests for conditional convergence as simple and general as those of §§ 167 et seq. It is naturally a much more difficult matter to formulate tests of convergence for series whose convergence, as is shown by equation (1) above, depends essentially on the cancelling of the positive by the negative terms. In the first instance there are no comparison tests for convergence of conditionally convergent series.

For suppose we wish to infer the convergence of vn from that of un. We have to compare

v0 + v1 + + vn,u0 + u1 + + un.
If every u and every v were positive, and every v less than the corresponding u, we could at once infer that
v0 + v1 + + vn < u0 + + un,
and so that vn is convergent. If the u’s only were positive and every v numerically less than the corresponding u, we could infer that
v0 + v1 + + vn < u0 + + un,
and so that vn is absolutely convergent. But in the general case, when the u’s and v’s are both unrestricted as to sign, all that we can infer is that
v0 + v1 + + vn < u0 + + un.
This would enable us to infer the absolute convergence of vn from the absolute convergence of  un; but if un is only conditionally convergent we can draw no inference at all.

Example. We shall see shortly that the series 1 1 2 + 1 3 1 4 + is convergent. But the series 1 2 + 1 3 + 1 4 + 1 5 + is divergent, although each of its terms is numerically less than the corresponding term of the former series.

It is therefore only natural that such tests as we can obtain should be of a much more special character than those given in the early part of this chapter.

188. Alternating Series. The simplest and most common conditionally convergent series are what is known as alternating series, series whose terms are alternately positive and negative. The convergence of the most important series of this type is established by the following theorem.

If φ(n) is a positive function of n which tends steadily to zero as n , then the series

φ(0) φ(1) + φ(2)
is convergent, and its sum lies between φ(0) and φ(0) φ(1).

Let us write φ0, φ1, … for φ(0), φ(1), …; and let

sn = φ0 φ1 + φ2 + (1)nφ n.
Then
s2n+1 s2n1 = φ2n φ2n+1 0,s2n s2n2 = (φ2n1 φ2n) 0.
Hence s0, s2, s4, …, s2n, … is a decreasing sequence, and therefore tends to a limit or to  , and s1, s3, s5, …, s2n+1, … is an increasing sequence, and therefore tends to a limit or to . But lim(s2n+1 s2n) = lim(1)2n+1φ 2n+1 = 0, from which it follows that both sequences must tend to limits, and that the two limits must be the same. That is to say, the sequence s0, s1, …, sn, … tends to a limit. Since s0 = φ0, s1 = φ0 φ1, it is clear that this limit lies between φ0 and φ0 φ1.

Examples LXXVIII. 1. The series

1 1 2 + 1 3 1 4 + ,1 1 2 + 1 3 1 4 + , (1)n (n + a),(1)n n + a , (1)n (n + a), (1)n (n + a)2,

where a > 0, are conditionally convergent.

2. The series (1)n(n + a)s, where a > 0, is absolutely convergent if s > 1, conditionally convergent if 0 < s 1, and oscillatory if s 0.

3. The sum of the series of §188 lies between sn and sn+1 for all values of n; and the error committed by taking the sum of the first n terms instead of the sum of the whole series is numerically not greater than the modulus of the (n + 1)th term.

4. Consider the series

(1)n n + (1)n,
which we suppose to begin with the term for which n = 2, to avoid any difficulty as to the definitions of the first few terms. This series may be written in the form
(1)n n + (1)n (1)n n + (1)n n
or
(1)n n 1 n + (1)nn = (ψn χn),
say. The series ψn is convergent; but χn is divergent, as all its terms are positive, and limnχn = 1. Hence the original series is divergent, although it is of the form φ2 φ3 + φ4 , where φn 0. This example shows that the condition that φn should tend steadily to zero is essential to the truth of the theorem. The reader will easily verify that 2n + 1 1 < 2n + 1, so that this condition is not satisfied.

5. If the conditions of §188 are satisfied except that φn tends steadily to a positive limit l, then the series (1)nφn oscillates finitely.

6. Alteration of the sum of a conditionally convergent series by rearrangement of the terms. Let s be the sum of the series 1 1 2 + 1 3 1 4 + , and s2n the sum of its first 2n terms, so that lims2n = s.

Now consider the series

1 + 1 3 1 2 + 1 5 + 1 7 1 4 + (1)

in which two positive terms are followed by one negative term, and let t3n denote the sum of the first 3n terms. Then

t3n = 1 + 1 3 + + 1 4n 1 1 2 1 4 1 2n = s2n + 1 2n + 1 + 1 2n + 3 + + 1 4n 1.

Now

lim 1 2n + 1 1 2n + 2 + 1 2n + 3 + 1 4n 1 1 4n = 0,
since the sum of the terms inside the bracket is clearly less than n/(2n + 1)(2n + 2); and
lim 1 2n + 2 + 1 2n + 4 + + 1 4n = 1 2 lim 1 n r=1n 1 1+(r/n) = 1 212dx x ,
by §§156 and 158. Hence
limt3n = s + 1 212dx x ,
and it follows that the sum of the series (1) is not s, but the right-hand side of the last equation. Later on we shall give the actual values of the sums of the two series: see §213 and Ch. IX, Misc. Ex. 19.

It can indeed be proved that a conditionally convergent series can always be so rearranged as to converge to any sum whatever, or to diverge to or to  . For a proof we may refer to Bromwich’s Infinite Series, p. 68.

7. The series

1 + 1 3 1 2 + 1 5 + 1 7 1 4 +
diverges to . [Here
t3n = s2n + 1 2n + 1 + 1 2n + 3 + + 1 4n 1 > s2n + n 4n 1,
where s2n = 1 1 2 + 1 2n, which tends to a limit as n .]

189. Abel’s and Dirichlet’s Tests of Convergence. A more general test, which includes the test of §188 as a particular test case, is the following.

Dirichlet’s Test. If φn satisfies the same conditions as in §188, and an is any series which converges or oscillates finitely, then the series

a0φ0 + a1φ1 + a2φ2 +
is convergent.

The reader will easily verify the identity

a0φ0 + a1φ1 + + anφn = s0(φ0 φ1) + s1(φ1 φ2) + + sn1(φn1 φn) + snφn,
where sn = a0 + a1 + + an. Now the series (φ0 φ1) + (φ1 φ2) + is convergent, since the sum to n terms is φ0 φn and limφn = 0; and all its terms are positive. Also since an, if not actually convergent, at any rate oscillates finitely, we can determine a constant K so that sν< K for all values of ν. Hence the series
sν(φν φν+1)
is absolutely convergent, and so
s0(φ0 φ1) + s1(φ1 φ2) + + sn1(φn1 φn)
tends to a limit as n . Also φn, and therefore snφn, tends to zero. And therefore
a0φ0 + a1φ1 + + anφn
tends to a limit, i.e. the series aνφν is convergent.

Abel’s Test. There is another test, due to Abel, which, though of less frequent application than Dirichlet’s, is sometimes useful.

Suppose that φn, as in Dirichlet’s Test, is a positive and decreasing function of n, but that its limit as n is not necessarily zero. Thus we postulate less about φn, but to make up for this we postulate more about an, viz. that it is convergent. Then we have the theorem: if φn is a positive and decreasing function of n, and an is convergent, then anφn is convergent.

For φn has a limit as n , say l: and lim(φn l) = 0. Hence, by Dirichlet’s Test, an(φn l) is convergent; and as an is convergent it follows that anφn is convergent.

This theorem may be stated as follows: a convergent series remains convergent if we multiply its terms by any sequence of positive and decreasing factors.

Examples LXXIX. 1. Dirichlet’s and Abel’s Tests may also be established by means of the general principle of convergence (§84). Let us suppose, for example, that the conditions of Abel’s Test are satisfied. We have identically

amφm + am+1φm+1 + + anφn = sm,m(φm φm+1) + sm,m+1(φm+1 φm+2) + + sm,n1(φn1 φn) + sm,nφn,(1)

where

sm,ν = am + am+1 + + aν.

The left-hand side of (1) therefore lies between hφm and Hφm, where h and H are the algebraically least and greatest of sm,m, sm,m+1, …, sm,n. But, given any positive number ε, we can choose m0 so that sm,ν< ε when m m0, and so

amφm + am+1φm+1 + + anφn< εφm εφ1
when n > m m0. Thus the series anφn is convergent.

2. The series cosnθ and sinnθ oscillate finitely when θ is not a multiple of π. For, if we denote the sums of the first n terms of the two series by sn and tn, and write z = Cisθ, so that z= 1 and z1, we have

sn + itn= 1 zn 1 z 1 + zn 1 z 2 1 z;
and so sn and tnare also not greater than 2/1 z. That the series are not actually convergent follows from the fact that their nth terms do not tend to zero (Exs. XXIV. 7, 8).

The sine series converges to zero if θ is a multiple of π. The cosine series oscillates finitely if θ is an odd multiple of π and diverges if θ is an even multiple of π.

It follows that if θn is a positive function of n which tends steadily to zero as n , then the series

φn cosnθ, φn sinnθ
are convergent, except perhaps the first series when θ is a multiple of 2π. In this case the first series reduces to  φn, which may or may not be convergent: the second series vanishes identically. If φn is convergent then both series are absolutely convergent (Ex. LXXVII. 4) for all values of θ, and the whole interest of the result lies in its application to the case in which φn is divergent. And in this case the series above written are conditionally and not absolutely convergent, as will be proved in Ex. LXXIX. 6. If we put θ = π in the cosine series we are led back to the result of §188, since cosnπ = (1)n.

3. The series ns cosnθ, ns sinnθ are convergent if s > 0, unless (in the case of the first series) θ is a multiple of 2π and 0 < s 1.

4. The series of Ex. 3 are in general absolutely convergent if s > 1, conditionally convergent if 0 < s 1, and oscillatory if s 0 (finitely if s = 0 and infinitely if s < 0). Mention any exceptional cases.

5. If anns is convergent or oscillates finitely, then annt is convergent when t > s.

6. If φn is a positive function of n which tends steadily to 0 as n , and φn is divergent, then the series φn cosnθ, φn sinnθ are not absolutely convergent, except the sine-series when θ is a multiple of π. [For suppose, e.g., that φncosnθis convergent. Since cos2nθ cosnθ, it follows that φn cos2nθ or

1 2 φn(1 + cos2nθ)
is convergent. But this is impossible, since φn is divergent and φn cos2nθ, by Dirichlet’s Test, convergent, unless θ is a multiple of π. And in this case it is obvious that φncosnθ is divergent. The reader should write out the corresponding argument for the sine-series, noting where it fails when θ is a multiple of π.]

190. Series of complex terms. So far we have confined ourselves to series all of whose terms are real. We shall now consider the series

un = (vn + iwn),
where vn and wn are real. The consideration of such series does not, of course, introduce anything really novel. The series is convergent if, and only if, the series
vn, wn
are separately convergent. There is however one class of such series so important as to require special treatment. Accordingly we give the following definition, which is an obvious extension of that of § 184.

DEFINITION. The series un, where un = vn + iwn, is said to be absolutely convergent if the series vn and wn are absolutely convergent.

THEOREM. The necessary and sufficient condition for the absolute convergence of  un is the convergence of un or vn2 + w n2.

For if un is absolutely convergent, then both of the series vn, wn are convergent, and so {vn + wn} is convergent: but

un = vn2 + w n2 v n + wn,
and therefore un is convergent. On the other hand
vnvn2 + w n2,w nvn2 + w n2,
so that vn and wn are convergent whenever un is convergent.

It is obvious that an absolutely convergent series is convergent, since its real and imaginary parts converge separately. And Dirichlet’s Theorem (§§ 169185) may be extended at once to absolutely convergent complex series by applying it to the separate series vn and  wn.

The convergence of an absolutely convergent series may also be deduced directly from the general principle of convergence (cf. Ex. LXXVII. 1). We leave this as an exercise to the reader.

191. Power Series. One of the most important parts of the theory of the ordinary functions which occur in elementary analysis (such as the sine and cosine, and the logarithm and exponential, which will be discussed in the next chapter) is that which is concerned with their expansion in series of the form anxn. Such a series is called a power series in x. We have already come across some cases of expansion in series of this kind in connection with Taylor’s and Maclaurin’s series (§ 148). There, however, we were concerned only with a real variable x. We shall now consider a few general properties of power series in z, where z is a complex variable.

A. A power series anzn may be convergent for all values of z, for a certain region of values, or for no values except z = 0.

It is sufficient to give an example of each possibility.

1. The series zn n! is convergent for all values of z. For if un = zn n! then

un+1/un= z/(n + 1) 0
as n , whatever value z may have. Hence, by d’Alembert’s Test, un is convergent for all values of z, and the original series is absolutely convergent for all values of z. We shall see later on that a power series, when convergent, is generally absolutely convergent.

2. The series n!zn is not convergent for any value of z except z = 0. For if un = n!zn then un+1/un= (n + 1)z, which tends to  with n, unless z = 0. Hence (cf. Exs. XXVII. 1, 2, 5) the modulus of the nth term tends to  with n; and so the series cannot converge, except when z = 0. It is obvious that any power series converges when z = 0.

3. The series zn is always convergent when z< 1, and never convergent when z1. This was proved in §88. Thus we have an actual example of each of the three possibilities.

192. B. If a power series anzn is convergent for a particular value of z, say z1 = r1(cos θ1 + i sin θ1), then it is absolutely convergent for all values of z such that z < r1.

For lim anz1n = 0, since anz1n is convergent, and therefore we can certainly find a constant K such that anz1n < K for all values of n. But, if z = r < r1, we have

anzn = a nz1nr r1 n < K r r1 n,
and the result follows at once by comparison with the convergent geometrical series (r/r1)n.

In other words, if the series converges at P then it converges absolutely at all points nearer to the origin than P.

Example. Show that the result is true even if the series oscillates finitely when z = z1. [If sn = a0 + a1z1 + + anz1n then we can find K so that sn< K for all values of n. But anz1n= sn sn1sn1+ sn< 2K, and the argument can be completed as before.]

193. The region of convergence of a power series. The circle of convergence. Let z = r be any point on the positive real axis. If the power series converges when z = r then it converges absolutely at all points inside the circle z = r. In particular it converges for all real values of z less than r.

Now let us divide the points r of the positive real axis into two classes, the class at which the series converges and the class at which it does not. The first class must contain at least the one point z = 0. The second class, on the other hand, need not exist, as the series may converge for all values of z. Suppose however that it does exist, and that the first class of points does include points besides z = 0. Then it is clear that every point of the first class lies to the left of every point of the second class. Hence there is a point, say the point z = R, which divides the two classes, and may itself belong to either one or the other. Then the series is absolutely convergent at all points inside the circle z = R.

For let P be any such point. We can draw a circle, whose centre is O and whose radius is


pict

Fig. 51.

less than R, so as to include P inside it. Let this circle cut OA in Q. Then the series is convergent at Q, and therefore, by Theorem B, absolutely convergent at P.

On the other hand the series cannot converge at any point Poutside the circle. For if it converged at P it would converge absolutely at all points nearer to O than P; and this is absurd, as it does not converge at any point between A and Q (Fig. 51).

So far we have excepted the cases in which the power series (1) does not converge at any point on the positive real axis except z = 0 or (2) converges at all points on the positive real axis. It is clear that in case (1) the power series converges nowhere except when z = 0, and that in case (2) it is absolutely convergent everywhere. Thus we obtain the following result: a power series either

(1) converges for z = 0 and for no other value of z; or

(2) converges absolutely for all values of z; or

(3) converges absolutely for all values of z within a certain circle of radius R, and does not converge for any value of z outside this circle.

In case (3) the circle is called the circle of convergence and its radius the radius of convergence of the power series.

It should be observed that this general result gives absolutely no information about the behaviour of the series on the circle of convergence. The examples which follow show that as a matter of fact there are very diverse possibilities as to this.

Examples LXXX. 1. The series 1 + az + a2z2 + , where a > 0, has a radius of convergence equal to 1/a. It does not converge anywhere on its circle of convergence, diverging when z = 1/a and oscillating finitely at all other points on the circle.

2. The series z 12 + z2 22 + z3 32 + has its radius of convergence equal to 1; it converges absolutely at all points on its circle of convergence.

3. More generally, if an+1/anλ, or an1/n λ, as n , then the series a0 + a1z + a2z2 + has 1/λ as its radius of convergence. In the first case

liman+1zn+1/a nzn= λz,
which is less or greater than unity according as z is less or greater than 1/λ, so that we can use d’Alembert’s Test (§168, 3). In the second case we can use Cauchy’s Test (§168, 2) similarly.

4. The logarithmic series. The series

z 1 2z2 + 1 3z3
is called (for reasons which will appear later) the ‘logarithmic’ series. It follows from Ex. 3 that its radius of convergence is unity.

When z is on the circle of convergence we may write z = cosθ + isinθ, and the series assumes the form

cosθ 1 2 cos2θ + 1 3 cos3θ + i(sinθ 1 2 sin2θ + 1 3 sin3θ ).

The real and imaginary parts are both convergent, though not absolutely convergent, unless θ is an odd multiple of π (Exs. LXXIX. 3, 4). If θ is an odd multiple of π then z = 1, and the series assumes the form 1 1 2 1 3 , and so diverges to  . Thus the logarithmic series converges at all points of its circle of convergence except the point z = 1.

5. The binomial series. Consider the series

1 + mz + m(m 1) 2! z2 + m(m 1)(m 2) 3! z3 + .
If m is a positive integer then the series terminates. In general
an+1 an = m n n + 1 1,
so that the radius of convergence is unity. We shall not discuss here the question of its convergence on the circle, which is a little more difficult.96

194. Uniqueness of a power series. If anzn is a power series which is convergent for some values of z at any rate besides z = 0, and f(z) is its sum, then it is easy to see that f(z) can be expressed in the form

a0 + a1z + a2z2 + + (a n + εz)zn,
where εz 0 as z0. For if μ is any number less than the radius of convergence of the series, and z< μ, then anμn < K, where K is a constant (cf. §192), and so

f(z) 0na νzν a n+1zn+1+ a n+2zn+2+ < K z μ n+1 1 + z μ + z2 μ2 + = Kzn+1 μn(μ z),

where K is a number independent of z. It follows from Ex. LV. 15 that if anzn = bnzn for all values of z whose modulus is less than some number μ, then an = bn for all values of n. This result is capable of considerable generalisations into which we cannot enter now. It shows that the same function f(z) cannot be represented by two different power series.

195. Multiplication of Series. We saw in § 170 that if un and vn are two convergent series of positive terms, then un × vn = wn, where

wn = u0vn + u1vn1 + + unv0.
We can now extend this result to all cases in which un and vn are absolutely convergent; for our proof was merely a simple application of Dirichlet’s Theorem, which we have already extended to all absolutely convergent series.

Examples LXXXI. 1. If z is less than the radius of convergence of either of the series anzn, bnzn, then the product of the two series is cnzn, where cn = a0bn + a1bn1 + + anb0.

2. If the radius of convergence of anzn is R, and f(z) is the sum of the series when z< R, and z is less than either R or unity, then f(z)/(1 z) = snzn, where sn = a0 + a1 + + an.

3. Prove, by squaring the series for 1/(1 z), that 1/(1 z)2 = 1 + 2z + 3z2 + if z< 1.

4. Prove similarly that 1/(1 z)3 = 1 + 3z + 6z2 + , the general term being 1 2(n + 1)(n + 2)zn.

5. The Binomial Theorem for a negative integral exponent. If z< 1, and m is a positive integer, then

1 (1 z)m = 1 + mz + m(m + 1) 1 2 z2 + + m(m + 1)(m + n 1) 1 2n zn + .

[Assume the truth of the theorem for all indices up to m. Then, by Ex. 2, 1/(1 z)m+1 = snzn, where

sn = 1 + m + m(m + 1) 1 2 + + m(m + 1)(m + n 1) 1 2n = (m + 1)(m + 2)(m + n) 1 2n ,

as is easily proved by induction.]

6. Prove by multiplication of series that if

f(m,z) = 1 + m 1 z + m 2 z2 + ,
and z< 1, then f(m,z)f(m,z) = f(m + m,z). [This equation forms the basis of Euler’s proof of the Binomial Theorem. The coefficient of zn in the product series is
m n + m 1 m n 1 + m 2 m n 2 + + m n 1 m 1 + m n .

This is a polynomial in m and m: but when m and m are positive integers this polynomial must reduce to m + m k in virtue of the Binomial Theorem for a positive integral exponent, and if two such polynomials are equal for all positive integral values of m and m then they must be equal identically.]

7. If f(z) = 1 + z + z2 2! + then f(z)f(z) = f(z + z). [For the series for f(z) is absolutely convergent for all values of z: and it is easy to see that if un = zn n! , vn = zn n! , then wn = (z + z)n n! .]

8. If

C(z) = 1 z2 2! + z4 4! ,S(z) = z z3 3! + z5 5! ,
then
C(z + z) = C(z)C(z) S(z)S(z),S(z + z) = S(z)C(z) + C(z)S(z),
and
{C(z)}2 + {S(z)}2 = 1.

9. Failure of the Multiplication Theorem. That the theorem is not always true when un and vn are not absolutely convergent may be seen by considering the case in which

un = vn = (1)n n + 1 .
Then
wn = (1)n r=0n 1 (r + 1)(n + 1 r).
But (r + 1)(n + 1 r) 1 2(n + 2), and so wn> (2n + 2)/(n + 2), which tends to 2; so that wn is certainly not convergent.

Miscellaneous Examples on Chapter VIII.

1. Discuss the convergence of the series nk{n + 1 2n + n 1}, where k is real.

(Math. Trip. 1890.)

2. Show that

nrΔk(ns),
where
Δun = un un+1,Δ2u n = Δ(Δun),
and so on, is convergent if and only if k > r + s + 1, except when s is a positive integer less than k, when every term of the series is zero.

[The result of Ch. VII, Misc. Ex. 11, shows that Δk(ns) is in general of order nsk.]

3. Show that

1 n2 + 9n + 5 (n + 1)(2n + 3)(2n + 5)(n + 4) = 5 36.
(Math. Trip. 1912.)

[Resolve the general term into partial fractions.]

4. Show that, if R(n) is any rational function of n, we can determine a polynomial P(n) and a constant A such that {R(n) P(n) (A/n)} is convergent. Consider in particular the cases in which R(n) is one of the functions 1/(an + b), (an2 + 2bn + c)/(αn2 + 2βn + γ).

5. Show that the series

1 1 1 + z + 1 2 1 2 + z + 1 3 1 3 + z +
is convergent provided only that z is not a negative integer.

6. Investigate the convergence or divergence of the series

sin a n,1 nsin a n, (1)n sin a n, 1 cos a n, (1)nn 1 cos a n,

where a is real.

7. Discuss the convergence of the series

11 + 1 2 + 1 3 + + 1 n sin(nθ + α) n ,
where θ and α are real.
(Math. Trip. 1989.)

8. Prove that the series

1 1 2 1 3 + 1 4 + 1 5 + 1 6 1 7 1 8 1 9 1 10 + ,
in which successive terms of the same sign form groups of 1, 2, 3, 4, … terms, is convergent; but that the corresponding series in which the groups contain 1, 2, 4, 8, … terms oscillates finitely.
(Math. Trip. 1908.)

9. If u1, u2, u3, … is a decreasing sequence of positive numbers whose limit is zero, then the series

u1 1 2(u1 + u2) + 1 3(u1 + u2 + u3) ,u1 1 3(u1 + u3) + 1 5(u1 + u3 + u5)
are convergent. [For if (u1 + u2 + + un)/n = vn then v1, v2, v3, … is also a decreasing sequence whose limit is zero (Ch. IV, Misc. Exs. 8, 27). This shows that the first series is convergent; the second we leave to the reader. In particular the series
1 1 2 1 + 1 2 + 1 3 1 + 1 2 + 1 3 ,1 1 3 1 + 1 3 + 1 5 1 + 1 3 + 1 5
are convergent.]

10. If u0 + u1 + u2 + is a divergent series of positive and decreasing terms, then

(u0 + u2 + + u2n)/(u1 + u3 + + u2n+1) 1.

11. Prove that if α > 0 then limp n=0(p + n)1α = 0.

12. Prove that limα0+α 1n1α = 1. [It follows from §174 that

0 < 11α + 21α + + (n 1)1α 1nx1αdx 1,
and it is easy to deduce that n1α lies between 1/α and (1/α) + 1.]

13. Find the sum of the series 1un, where

un = xn xn1 (xn + xn)(xn+1 + xn1) = 1 x 1 1 xn + xn 1 xn+1 + xn1 ,
for all real values of x for which the series is convergent.
(Math. Trip. 1901.)

[If x is not equal to unity then the series has the sum x/{(x 1)(x2 + 1)}. If x = 1 then un = 0 and the sum is 0. If x = 1 then un = 1 2(1)n+1 and the series oscillates finitely.]

14. Find the sums of the series

z 1 + z + 2z2 1 + z2 + 4z4 1 + z4 + , z 1 z2 + z2 1 z4 + z4 1 z8 +
(in which all the indices are powers of 2), whenever they are convergent.

[The first series converges only if z< 1, its sum then being z/(1 z); the second series converges to z/(1 z) if z< 1 and to 1/(1 z) if z> 1.]

15. If an1 for all values of n then the equation

0 = 1 + a1z + a2z2 +
cannot have a root whose modulus is less than 1 2, and the only case in which it can have a root whose modulus is equal to 1 2 is that in which an = Cis(nθ), when z = 1 2 Cis(θ) is a root.

16. Recurring Series. A power series anzn is said to be a recurring series if its coefficients satisfy a relation of the type

an + p1an1 + p2an2 + + pkank = 0, (1)

where n k and p1, p2, …, pk are independent of n. Any recurring series is the expansion of a rational function of z. To prove this we observe in the first place that the series is certainly convergent for values of z whose modulus is sufficiently small. For let G be the greater of the two numbers

1,p1+ p2+ + pk.
Then it follows from the equation (1) that anGαn, where αn is the modulus of the numerically greatest of the preceding coefficients; and from this that an< KGn, where K is independent of n. Thus the recurring series is certainly convergent for values of z whose modulus is less than 1/G.

But if we multiply the series f(z) = anzn by p1z, p2z2, …, pkzk, and add the results, we obtain a new series in which all the coefficients after the (k 1)th vanish in virtue of the relation (1), so that

(1 + p1z + p2z2 + + p kzk)f(z) = P 0 + P1z + + Pk1zk1,
where P0, P1, …, Pk1 are constants. The polynomial 1 + p1z + p2z2 + + pkzk is called the scale of relation of the series.

Conversely, it follows from the known results as to the expression of any rational function as the sum of a polynomial and certain partial fractions of the type A/(z a)p, and from the Binomial Theorem for a negative integral exponent, that any rational function whose denominator is not divisible by z can be expanded in a power series convergent for values of z whose modulus is sufficiently small, in fact if z< ρ, where ρ is the least of the moduli of the roots of the denominator (cf. Ch. IV, Misc. Exs. 18 et seq.). And it is easy to see, by reversing the argument above, that the series is a recurring series. Thus the necessary and sufficient condition that a power series should be a recurring series is that it should be the expansion of such a rational function of z.

17. Solution of Difference-Equations. A relation of the type of (1) in Ex. 16 is called a linear difference-equation in an with constant coefficients. Such equations may be solved by a method which will be sufficiently explained by an example. Suppose that the equation is

an an1 8an2 + 12an3 = 0.
Consider the recurring power series anzn. We find, as in Ex. 16, that its sum is
a0 + (a1 a0)z + (a2 a1 8a0)z2 1 z 8z2 + 12z3 = A1 1 2z + A2 (1 2z)2 + B 1 + 3z,
where A1A2, and B are numbers easily expressible in terms of a0a1, and a2. Expanding each fraction separately we see that the coefficient of zn is
an = 2n{A 1 + (n + 1)A2}+ (3)nB.
The values of A1A2B depend upon the first three coefficients a0a1a2, which may of course be chosen arbitrarily.

18. The solution of the difference-equation un 2cosθun1 + un2 = 0 is un = Acosnθ + Bsinnθ, where A and B are arbitrary constants.

19. If un is a polynomial in n of degree k, then unzn is a recurring series whose scale of relation is (1 z)k+1.

(Math. Trip. 1904.)

20. Expand 9/{(z 1)(z + 2)2}in ascending powers of z.

(Math. Trip. 1913.)

21. Prove that if f(n) is the coefficient of zn in the expansion of z/(1 + z + z2) in powers of z, then

(1) f(n) + f(n 1) + f(n 2) = 0,(2) f(n) = (ω3n ω 32n)/(ω 3 ω32),
where ω3 is a complex cube root of unity. Deduce that f(n) is equal to 0 or 1 or  1 according as n is of the form 3k or 3k + 1 or 3k + 2, and verify this by means of the identity z/(1 + z + z2) = z(1 z)/(1 z3).

22. A player tossing a coin is to score one point for every head he turns up and two for every tail, and is to play on until his score reaches or passes a total n. Show that his chance of making exactly the total n is 1 3{2 + (1 2)n}.

(Math. Trip. 1898.)

[If pn is the probability then pn = 1 2(pn1 + pn2). Also p0 = 1, p1 = 1 2.]

23. Prove that

1 a + 1 + 1 a + 2 + + 1 a + n = n 1 1 a + 1 n 2 1! (a + 1)(a + 2) +
if n is a positive integer and a is not one of the numbers 1, 2, …,  n.

[This follows from splitting up each term on the right-hand side into partial fractions. When a > 1, the result may be deduced very simply from the equation

01xa1 xn 1 x dx =01(1 x)a{1 (1 x)n}dx x
by expanding (1 xn)/(1 x) and 1 (1 x)n in powers of x and integrating each term separately. The result, being merely an algebraical identity, must be true for all values of a save 1, 2, …,  n.]

24. Prove by multiplication of series that

0zn n! 1(1)n1zn n n! = 11 + 1 2 + 1 3 + + 1 n zn n! .

[The coefficient of zn will be found to be

1 n! n 1 1 2 n 2 + 1 3 n 3 .
Now use Ex. 23, taking a = 0.]

25. If An A and Bn B as n , then

(A1Bn + A2Bn1 + + AnB1)/n AB.

[Let An = A + εn. Then the expression given is equal to

AB1 + B2 + + Bn n + ε1Bn + ε2Bn1 + + εnB1 n .

The first term tends to AB (Ch. IV, Misc. Ex. 27). The modulus of the second is less than β{ε1+ ε2+ + εn}/n, where β is any number greater than the greatest value of Bν: and this expression tends to zero.]

26. Prove that if cn = a1bn + a2bn1 + + anb1 and

An = a1 + a2 + + an,Bn = b1 + b2 + + bn,Cn = c1 + c2 + + cn,
then
Cn = a1Bn + a2Bn1 + + anB1 = b1An + b2An1 + + bnA1
and
C1 + C2 + + Cn = A1Bn + A2Bn1 + + AnB1.

Hence prove that if the series an, bn are convergent and have the sums AB, so that An A, Bn B, then

(C1 + C2 + + Cn)/n AB.
Deduce that if cn is convergent then its sum is AB. This result is known as Abel’s Theorem on the multiplication of Series. We have already seen that we can multiply the series an, bn in this way if both series are absolutely convergent: Abel’s Theorem shows that we can do so even if one or both are not absolutely convergent, provided only that the product series is convergent.

27. Prove that

1 2 1 1 2 + 1 3 2 = 1 2 1 3 1 + 1 2 + 1 4 1 + 1 2 + 1 3 , 1 2 1 1 3 + 1 5 2 = 1 2 1 4 1 + 1 3 + 1 6 1 + 1 3 + 1 5 .

[Use Ex. 9 to establish the convergence of the series.]

28. For what values of m and n is the integral 0π sinmx(1 cosx)ndx convergent? [If m + 1 and m + 2n + 1 are positive.]

29. Prove that if a > 1 then

11 dx (a x)1 x2 = π a2 1.

30. Establish the formulae

0F{x2 + 1 + x}dx = 1 21 1 + 1 y2 F(y)dy, 0F{x2 + 1 x}dx = 1 201 1 + 1 y2 F(y)dy.

In particular, prove that if n > 1 then

0 dx {x2 + 1 + x}n =0{x2 + 1 x}ndx = n n2 1.

[In this and the succeeding examples it is of course supposed that the arbitrary functions which occur are such that the integrals considered have a meaning in accordance with the definitions of §§177 et seq.]

31. Show that if 2y = ax (b/x), where a and b are positive, then y increases steadily from to as x increases from 0 to . Hence show that

0f 1 2 ax + b x dx = 1 af{y2 + ab}1 + y y2 + abdy = 2 a0f{y2 + ab}dy.

32. Show that if 2y = ax + (b/x), where a and b are positive, then two values of x correspond to any value of y greater than ab. Denoting the greater of these by x1 and the less by x2, show that, as y increases from ab towards , x1 increases from b/a towards , and x2 decreases from b/a to 0. Hence show that

b/af(y)dx 1 = 1 aabf(y) y y2 ab + 1dy, 0b/af(y)dx 2 = 1 aabf(y) y y2 ab 1dy,

and that

0f 1 2 ax + b x dx = 2 aab yf(y) y2 abdy = 2 a0f{z2 + ab}dz.

33. Prove the formula

0πf(sec 1 2x + tan 1 2x) dx sin x =0πf(cosecx) dx sin x.

34. If a and b are positive, then

0 dx (x2 + a2)(x2 + b2) = π 2ab(a + b),0 x2dx (x2 + a2)(x2 + b2) = π 2(a + b).
Deduce that if αβ, and γ are positive, and β2 αγ, then
0 dx αx4 + 2βx2 + γ = π 22γA,0 x2dx αx4 + 2βx2 + γ = π 22αA,
where A = β + αγ. Also deduce the last result from Ex. 31, by putting f(y) = 1/(c2 + y2). The last two results remain true when β2 < αγ, but their proof is then not quite so simple.

35. Prove that if b is positive then

0 x2dx (x2 a2)2 + b2x2 = π 2b,0 x4dx {(x2 a2)2 + b2x2}2 = π 4b3.

36. Extend Schwarz’s inequality (Ch. VII, Misc. Ex. 42) to infinite integrals of the first and second kinds.

37. Prove that if φ(x) is the function considered at the end of §178 then

0φ(x)dx = 0 1 (n + 1)2.

38. Prove that

1dx 1 x y (x + y)3dy = 1,1dy 1 x y (x + y)3dx = 1; 1dx 1 x2 y2 (x2 + y2)2dy = 1 4π,1dy 1 x2y2 (x2+y2)2dx = 1 4π.

Establish similar results in which the limits of integration are 0 and 1.

(Math. Trip. 1913.)