2.3 Series of Real Numbers
An infinite sum is the limit of its partial sums. We build the toolkit for deciding when that limit exists — comparison, condensation, integral, root, ratio, Raabe, Leibniz, Dirichlet and Abel — and see how fragile a conditionally convergent sum is.
A series $a_1+a_2+a_3+\cdots$ looks like an addition problem, but no one can add infinitely many numbers. What we can do is add the first $n$ of them and ask whether these running totals, the partial sums, form a convergent sequence. So everything in this section rests on §§2.1–2.2: a series converges when its sequence of partial sums converges, and the Cauchy criterion, the monotone convergence theorem and $\limsup$ do the real work. Series appear everywhere in analysis. The number $e$, the functions $e^x$, $\sin x$ and $\log(1+x)$, the solutions of differential equations and the Fourier expansion of a signal are all defined or computed by infinite series. Whether a numerical method converges, and how fast, is a question about the tail of a series.
The central difficulty is that we can rarely find the sum. Instead we compare: with geometric series (the ratio and root tests), with the $p$-series $\sum n^{-p}$ (comparison, condensation and integral tests, Raabe's test), and with the cancellation in alternating sums (the Leibniz, Dirichlet and Abel tests). The distinction between absolute and conditional convergence then turns out to be decisive. Absolutely convergent series behave like finite sums: they can be rearranged and multiplied freely. Conditionally convergent ones cannot, and Riemann's theorem says that by reordering their terms we can make them add up to any number we like. CSIR NET asks about series in almost every paper: the one convergent series among four, the values of $p$ or $x$ for which a series converges, and "which of the following are true" statements whose false options are the classic counterexamples collected in this section.
Series and Partial Sums; the Cauchy Criterion
The definition turns every question about series into a question about the sequence $(s_n)$. Conversely every sequence $(x_n)$ is the sequence of partial sums of the series $x_1+\sum_{n\ge2}(x_n-x_{n-1})$, so series and sequences are two languages for the same thing. The first theorem translates the results of §2.1.
- Linearity. If $\sum a_n=A$ and $\sum b_n=B$, then $\sum(\alpha a_n+\beta b_n)=\alpha A+\beta B$ for all $\alpha,\beta\in\mathbb{R}$.
- Cauchy criterion. $\sum a_n$ converges if and only if for every $\varepsilon\gt0$ there is $N$ such that $$|a_{m+1}+a_{m+2}+\cdots+a_n|\lt\varepsilon\qquad\text{for all }n\gt m\ge N .$$
- $n$-th term test. If $\sum a_n$ converges, then $a_n\to0$. Equivalently: if $a_n\not\to0$, the series diverges.
- Tails. Changing, adding or deleting finitely many terms does not affect convergence (it may change the sum). If $\sum a_n$ converges, its remainders satisfy $r_N\to0$.
- Grouping. If $\sum a_n=s$ and the terms are grouped into consecutive finite blocks without changing their order, the series of block sums also converges to $s$. Conversely, if the series of block sums converges to $s$, then $\sum a_n=s$ provided either $a_n\to0$ and the blocks have bounded length, or all terms inside each block have the same sign.
Proof. (1) The partial sums of $\sum(\alpha a_n+\beta b_n)$ are $\alpha s_n(a)+\beta s_n(b)$; apply the algebra of limits (§2.1, Theorem 2.2). (2) Since $s_n-s_m=a_{m+1}+\cdots+a_n$, this is exactly the statement that $(s_n)$ is a Cauchy sequence, and a real sequence converges if and only if it is Cauchy (§2.1, Theorem 2.8). (3) $a_n=s_n-s_{n-1}\to s-s=0$. (4) If two series differ in only the first $N$ terms, then for $n\ge N$ their partial sums differ by a fixed constant, so both converge or both diverge. And $r_N=s-s_N\to s-s=0$.
(5) If the blocks end at indices $n_1\lt n_2\lt\cdots$, the partial sums of the grouped series are $s_{n_1},s_{n_2},\dots$, which form a subsequence of $(s_n)$. If $s_n\to s$, the subsequence converges to $s$ (§2.2, Theorem 2.11). For the converse, suppose $s_{n_k}\to s$. Every $n\gt n_1$ satisfies $n_k\le n\lt n_{k+1}$ for exactly one $k$. If the blocks have length at most $L$, then $|s_n-s_{n_k}|\le L\max_{j\gt n_k}|a_j|$, and the right side tends to $0$ as $k\to\infty$ when $a_n\to0$. If the terms of each block have one sign, then $s_n$ lies between $s_{n_k}$ and $s_{n_{k+1}}$. In both cases $s_n\to s$. ∎
The converse in (5) really needs a hypothesis. The series $1-1+1-1+\cdots$ diverges, since its terms do not tend to $0$, yet the grouping $(1-1)+(1-1)+\cdots$ converges to $0$ and the grouping $1+(-1+1)+(-1+1)+\cdots$ converges to $1$. You may insert brackets into a convergent series, but you may not remove them freely.
Proof. (a) For $r\ne1$, $s_n=1+r+\cdots+r^{n}=\dfrac{1-r^{n+1}}{1-r}$. If $|r|\lt1$, then $r^{n+1}\to0$ (§2.1, Theorem 2.5) and $s_n\to\frac1{1-r}$. If $|r|\ge1$, then $|r^n|\ge1$, so the terms do not tend to $0$ and the series diverges by Theorem 2.21(3). (b) The partial sums collapse: $s_n=(b_1-b_2)+\cdots+(b_n-b_{n+1})$ equals $b_1-b_{n+1}$. ∎
Proof. $s_{n+1}-s_n=a_{n+1}\ge0$, so $(s_n)$ is increasing. By the monotone convergence theorem (§2.1, Theorem 2.6) it converges if it is bounded, and it tends to $+\infty$ if it is not. ∎
Almost every test below is a way of proving that partial sums are bounded without computing them. That is why the theory of series with non-negative terms comes first, and why the general case is then handled through absolute values.
Comparison, Condensation and the Integral Test
For a series of non-negative terms, convergence means bounded partial sums, so it is enough to bound the terms by those of a series we already know converges. The two families we compare with are the geometric series and the $p$-series.
(b) Limit comparison test. Let $a_n\ge0$, $b_n\gt0$ and $\dfrac{a_n}{b_n}\to L\in[0,\infty]$.
- If $0\lt L\lt\infty$, then $\sum a_n$ and $\sum b_n$ both converge or both diverge.
- If $L=0$ and $\sum b_n$ converges, then $\sum a_n$ converges.
- If $L=\infty$ and $\sum b_n$ diverges, then $\sum a_n$ diverges.
Proof. (a) By Theorem 2.21(4) we may assume $N=1$. Then $s_n(a)\le s_n(b)\le\sum b_k$, so the partial sums of $\sum a_n$ are bounded and Theorem 2.23 applies. The second statement is the contrapositive of the first. (b) If $0\lt L\lt\infty$, then for $n$ large $\frac L2b_n\le a_n\le2Lb_n$, and (a) gives both directions. If $L=0$, then eventually $a_n\le b_n$. If $L=\infty$, then eventually $a_n\ge b_n$. ∎
The hypothesis $a_n\ge0$ cannot be dropped from (b). With $b_n=\frac{(-1)^n}{\sqrt n}$ and $a_n=\frac{(-1)^n}{\sqrt n}+\frac1n$ we have $a_n/b_n=1+\frac{(-1)^n}{\sqrt n}\to1$, but $\sum b_n$ converges (Leibniz test below) while $\sum a_n$ diverges (it is $\sum b_n$ plus the harmonic series).
Proof. Group the terms in blocks of lengths $1,2,4,\dots$. The block $a_{2^j}+\cdots+a_{2^{j+1}-1}$ has $2^j$ terms, each at most $a_{2^j}$ because $(a_n)$ decreases. Hence $$s_{2^{k+1}-1}=\sum_{j=0}^{k}\big(a_{2^j}+\cdots+a_{2^{j+1}-1}\big)\le\sum_{j=0}^k2^ja_{2^j}.$$ Grouping instead as $a_1+a_2+(a_3+a_4)+(a_5+\cdots+a_8)+\cdots$, the block $a_{2^{j-1}+1}+\cdots+a_{2^j}$ has $2^{j-1}$ terms, each at least $a_{2^j}$, so $$s_{2^k}\ge a_1+\sum_{j=1}^k2^{j-1}a_{2^j}\ge\frac12\sum_{j=0}^k2^ja_{2^j}.$$ The first inequality shows that if the condensed series converges, the partial sums $s_n\le s_{2^{k+1}-1}$ are bounded. The second shows that if $\sum a_n$ converges, the partial sums of the condensed series are bounded by $2\sum a_n$. Theorem 2.23 finishes both directions. ∎
(b) $\displaystyle\sum_{n=2}^\infty\frac1{n(\log n)^p}$ converges if and only if $p\gt1$.
(c) $\displaystyle\sum_{n=3}^\infty\frac1{n\log n\,(\log\log n)^p}$ converges if and only if $p\gt1$.
In particular $\sum\frac1n$, $\sum\frac1{n\log n}$ and $\sum\frac1{n\log n\log\log n}$ all diverge, each more slowly than the one before.
Proof. (a) If $p\le0$, then $\frac1{n^p}\ge1$ and the terms do not tend to $0$. If $p\gt0$, the terms decrease, and the condensed series is $$\sum_k2^k\cdot2^{-kp}=\sum_k\big(2^{1-p}\big)^k,$$ a geometric series that converges if and only if $2^{1-p}\lt1$, that is, $p\gt1$ (Theorem 2.22(a)). (b) For $p\gt0$ the terms decrease, and $2^ka_{2^k}=\dfrac{2^k}{2^k(k\log2)^p}=\dfrac1{(\log2)^p}\cdot\dfrac1{k^p}$, so the condensed series is a constant times the $p$-series of (a). For $p\le0$ the terms are at least $\frac1n$ for $n\ge3$, so the series diverges. (c) Condensing once more turns (c) into a constant multiple of (b), up to finitely many terms. ∎
- $\sum f(n)$ converges if and only if the improper integral $\int_1^\infty f=\lim_{b\to\infty}\int_1^bf$ is finite (§4.2).
- $d_n=s_n-\int_1^nf$ is decreasing and $0\le d_n\le f(1)$, so $d_n$ converges to a limit in $[0,f(1)]$.
- If the series converges, its remainders satisfy $$\int_{N+1}^\infty f\le\sum_{n=N+1}^\infty f(n)\le\int_N^\infty f .$$
Proof. For $k\le x\le k+1$ we have $f(k+1)\le f(x)\le f(k)$. Integrating over $[k,k+1]$ gives $$f(k+1)\le\int_k^{k+1}f\le f(k).$$ Summing over $k=1,\dots,n-1$ gives $s_n-f(1)\le\int_1^nf\le s_{n-1}$. (1) If the integral is finite, $s_n\le f(1)+\int_1^\infty f$ is bounded. If it is infinite, $s_{n-1}\ge\int_1^nf\to\infty$. (2) Since $d_n-d_{n+1}=\int_n^{n+1}f-f(n+1)\ge0$, the sequence decreases. Also $d_n=f(n)+\sum_{k=1}^{n-1}\big(f(k)-\int_k^{k+1}f\big)\ge f(n)\ge0$, and $d_1=f(1)$. (3) Summing $f(k+1)\le\int_k^{k+1}f\le f(k)$ over $k\ge N$ gives $\sum_{n\gt N}f(n)\le\int_N^\infty f$. Summing over $k\ge N+1$ gives $\int_{N+1}^\infty f\le\sum_{n\gt N}f(n)$. ∎
The Root, Ratio and Raabe Tests
The root and ratio tests compare a series with a geometric series. The root test uses $\limsup$ (§2.2), so it never needs a limit to exist.
- If $\rho\lt1$, then $\sum a_n$ converges absolutely (that is, $\sum|a_n|$ converges).
- If $\rho\gt1$, then $|a_n|\gt1$ for infinitely many $n$, so $a_n\not\to0$ and $\sum a_n$ diverges.
- If $\rho=1$, no conclusion is possible: $\sum\frac1n$ diverges and $\sum\frac1{n^2}$ converges, and both have $\rho=1$.
Proof. If $\rho\lt1$, fix $r$ with $\rho\lt r\lt1$. By Theorem 2.15(2) of §2.2, $|a_n|^{1/n}\lt r$ for all large $n$, so $|a_n|\lt r^n$ eventually. Compare with the geometric series $\sum r^n$. If $\rho\gt1$, then $|a_n|^{1/n}\gt1$ for infinitely many $n$ (again Theorem 2.15(2), or directly from the definition when $\rho=+\infty$), so infinitely many terms satisfy $|a_n|\gt1$. Finally $n^{-1/n}\to1$ and $n^{-2/n}\to1$, which gives the two examples. ∎
- If $\limsup\left|\dfrac{a_{n+1}}{a_n}\right|\lt1$, then $\sum a_n$ converges absolutely.
- If $\left|\dfrac{a_{n+1}}{a_n}\right|\ge1$ for all $n\ge N$ (in particular, if $\liminf\left|\dfrac{a_{n+1}}{a_n}\right|\gt1$), then $\sum a_n$ diverges.
- Otherwise ($\liminf\le1\le\limsup$) no conclusion is possible. In particular $\limsup\left|\frac{a_{n+1}}{a_n}\right|\gt1$ does not imply divergence.
Proof. If $\limsup|a_{n+1}/a_n|\lt1$, then $\limsup|a_n|^{1/n}\lt1$ by the chain of inequalities in Theorem 2.18 (§2.2), and the root test applies. (Directly: for $r$ strictly between the $\limsup$ and $1$, $|a_n|\le|a_N|r^{n-N}$ for $n\ge N$.) If $|a_{n+1}|\ge|a_n|$ for $n\ge N$, then $|a_n|\ge|a_N|\gt0$ for all $n\ge N$, so $a_n\not\to0$. Example 4(d) below shows the last claim. ∎
By Theorem 2.18, $\liminf\frac{|a_{n+1}|}{|a_n|}\le\liminf|a_n|^{1/n}$ and $\limsup|a_n|^{1/n}\le\limsup\frac{|a_{n+1}|}{|a_n|}$. Hence whenever the ratio test decides, the root test decides too, and the root test sometimes decides when the ratio test cannot. The ratio test is often easier to compute, especially with factorials. Both fail for every series whose terms decay like a power of $n$, because then both limits equal $1$. For these, Raabe's test measures the ratio more finely.
- If there are $r\gt1$ and $N$ with $R_n\ge r$ for all $n\ge N$, then $\sum a_n$ converges.
- If $R_n\le1$ for all $n\ge N$, then $\sum a_n$ diverges.
Proof. Suppose $R_n\ge r\gt1$ for $n\ge N$. Then $na_n-na_{n+1}\ge ra_{n+1}$, that is, $$na_n-(n+1)a_{n+1}\ge(r-1)a_{n+1}\gt0\qquad(n\ge N).$$ Summing from $n=N$ to $m-1$, the left side telescopes: $(r-1)\sum_{n=N+1}^{m}a_n\le Na_N-ma_m\le Na_N$. The partial sums are bounded, so the series converges (Theorem 2.23). Now suppose $R_n\le1$ for $n\ge N$. Then $na_n-na_{n+1}\le a_{n+1}$, so $na_n\le(n+1)a_{n+1}$. The sequence $(na_n)_{n\ge N}$ is increasing, so $a_n\ge\frac cn$ with $c=Na_N\gt0$, and the series diverges by comparison with the harmonic series. For $R=1$: both $\frac1{n\log n}$ (divergent) and $\frac1{n(\log n)^2}$ (convergent) have $R_n\to1$. ∎
A sharper version, Gauss's test, says: if $\dfrac{a_n}{a_{n+1}}=1+\dfrac\mu n+O\!\left(\dfrac1{n^2}\right)$, then $\sum a_n$ converges if and only if $\mu\gt1$ (statement only). The case $\mu=1$ diverges.
- Do the terms tend to $0$? If not, the series diverges, and you are finished.
- Non-negative terms built from powers, roots and logarithms: find the dominant behaviour $a_n\approx\frac{C}{n^p(\log n)^q}$ and use limit comparison with the $p$- and log-scale of Theorem 2.26. Taylor expansions ($\sin\frac1n$, $1-\cos\frac1n$, $e^{1/n}-1$, $n^{1/n}-1$) are the usual source of the dominant term.
- Factorials, $n$-th powers, $c^n$: ratio test for products and factorials, root test for expressions raised to the power $n$ or $n^2$. If the limit is $1$, switch to Raabe's test or to Stirling-type estimates.
- Decreasing positive terms with logarithms: condensation or the integral test.
- Signs that change: first test absolute convergence. If that fails, try Leibniz (alternating, decreasing to $0$), then Dirichlet (bounded partial sums of the oscillating factor times a monotone null factor), then Abel. If none applies, split the term into a convergent part plus a remainder (Example 5(c)).
- Depends on a parameter $x$ or $p$: find the critical value with the ratio or root test, then examine the boundary values separately. They almost always need a different test.
Absolute and Conditional Convergence; Alternating Series; Dirichlet and Abel
(b) $\sum a_n$ converges absolutely if and only if both $\sum a_n^+$ and $\sum a_n^-$ converge, and then $\sum a_n=\sum a_n^+-\sum a_n^-$.
(c) If $\sum a_n$ converges conditionally, then $\sum a_n^+=\sum a_n^-=+\infty$: the positive terms alone and the negative terms alone both diverge.
Proof. (a) By the triangle inequality, $|a_{m+1}+\cdots+a_n|$ is at most $|a_{m+1}|+\cdots+|a_n|$. The right side is small for large $m$ by the Cauchy criterion for $\sum|a_n|$, so the left side is too, and $\sum a_n$ satisfies the Cauchy criterion (Theorem 2.21(2)). Letting $n\to\infty$ in $|s_n|\le\sum_{k\le n}|a_k|$ gives the inequality. (b) If $\sum|a_n|$ converges, the comparisons $0\le a_n^\pm\le|a_n|$ give convergence of both parts. Conversely $|a_n|=a_n^++a_n^-$, and linearity applies. (c) If $\sum a_n^+$ converged, then $\sum a_n^-=\sum(a_n^+-a_n)$ would converge by linearity, and by (b) $\sum a_n$ would converge absolutely. The same argument works with $+$ and $-$ exchanged. Both series have non-negative terms, so they diverge to $+\infty$ (Theorem 2.23). ∎
Proof. Since $s_{2n+2}-s_{2n}=a_{2n+1}-a_{2n+2}\ge0$ and $s_{2n+1}-s_{2n-1}=-a_{2n}+a_{2n+1}\le0$, the even partial sums increase and the odd ones decrease. Also $s_{2n+1}-s_{2n}=a_{2n+1}\ge0$, so $$s_2\le s_4\le\cdots\le s_{2n}\le s_{2n+1}\le\cdots\le s_3\le s_1 .$$ Both monotone sequences are bounded, so they converge (§2.1, Theorem 2.6). Their difference $a_{2n+1}$ tends to $0$, so they have the same limit $s$, and $s_n\to s$ (§2.2, Theorem 2.11(d)). From the display, $s$ lies between $s_n$ and $s_{n+1}$ for every $n$. Hence $|s-s_n|\le|s_{n+1}-s_n|=a_{n+1}$, and $s-s_n$ has the sign of $s_{n+1}-s_n=(-1)^na_{n+1}$. ∎
So $1-\frac12+\frac13-\cdots$ converges (to $\log2$, Exercise III.7) but not absolutely. Leibniz's test is the special case $a_n=(-1)^{n+1}$ of a more flexible principle: bounded partial sums times a monotone null sequence. The tool behind it is a discrete form of integration by parts.
Proof. Substitute $a_k=A_k-A_{k-1}$ and shift the index in the second sum: $$\sum_{k=m}^n(A_k-A_{k-1})b_k=\sum_{k=m}^nA_kb_k-\sum_{k=m-1}^{n-1}A_kb_{k+1}.$$ Separating the term $k=n$ from the first sum and the term $k=m-1$ from the second leaves $A_nb_n-A_{m-1}b_m+\sum_{k=m}^{n-1}A_k(b_k-b_{k+1})$. ∎
(b) Abel's test. If $\sum a_n$ converges and $(b_n)$ is monotone and bounded, then $\sum a_nb_n$ converges.
Leibniz's test is (a) with $a_n=(-1)^{n+1}$, whose partial sums are $1,0,1,0,\dots$
Proof. (a) Since $(b_n)$ is monotone with limit $0$, $(|b_n|)$ decreases and the differences $b_k-b_{k+1}$ all have the same sign. So $\sum_{k=m}^{n-1}|b_k-b_{k+1}|=|b_m-b_n|$. By Theorem 2.33, $$\Big|\sum_{k=m}^na_kb_k\Big|\le M|b_n|+M|b_m|+M|b_m-b_n|\le4M|b_m| .$$ The right side tends to $0$ as $m\to\infty$, so the Cauchy criterion holds. (b) $(b_n)$ converges to some $b$ by the monotone convergence theorem, and $b_n-b$ is monotone with limit $0$. The partial sums of the convergent series $\sum a_n$ are bounded, so $\sum a_n(b_n-b)$ converges by (a). Then $\sum a_nb_n=\sum a_n(b_n-b)+b\sum a_n$ converges by linearity. ∎
The standard application uses the two trigonometric sums. For $x\notin2\pi\mathbb{Z}$, the product formulas $2\sin\frac x2\sin kx=\cos(k-\frac12)x-\cos(k+\frac12)x$ and $2\sin\frac x2\cos kx=\sin(k+\frac12)x-\sin(k-\frac12)x$ make the sums telescope: $$\Big|\sum_{k=1}^n\sin kx\Big|\le\frac1{|\sin(x/2)|},\qquad\Big|\sum_{k=1}^n\cos kx\Big|\le\frac1{|\sin(x/2)|}.$$
Rearrangements and Products of Series
Finite sums can be added in any order and multiplied term by term. For infinite series both operations are a genuine issue, and the answer depends on whether the convergence is absolute.
Proof. Let $t_k=a_{\sigma(1)}+\cdots+a_{\sigma(k)}$ and $\varepsilon\gt0$. Choose $N$ with $\sum_{n\gt N}|a_n|\lt\varepsilon$ (Theorem 2.21(4)), so that also $|s-s_N|\lt\varepsilon$. Since $\sigma$ is onto, there is $K$ such that $\{\sigma(1),\dots,\sigma(K)\}\supseteq\{1,\dots,N\}$. For $k\ge K$, the difference $t_k-s_N$ is a finite sum of terms $a_n$ with distinct indices $n\gt N$, so $|t_k-s_N|\le\sum_{n\gt N}|a_n|\lt\varepsilon$. Hence $|t_k-s|\lt2\varepsilon$ for all $k\ge K$, and $t_k\to s$. Applying the same argument to $\sum|a_n|$ shows that the rearrangement converges absolutely. For non-negative terms, every partial sum of either series is at most the sum of the other, so the two sums coincide. ∎
Proof sketch. Let $p_1,p_2,\dots$ be the positive terms of $\sum a_n$ and $-q_1,-q_2,\dots$ the negative terms, each list in its original order. Zero terms can be slotted in anywhere without effect. Step 1: by Theorem 2.31(c), $\sum p_i=\sum q_j=+\infty$. Step 2: $p_i\to0$ and $q_j\to0$, because $a_n\to0$. Step 3 (greedy construction for real $s$): take $p_1,p_2,\dots$ until the running sum first exceeds $s$. Then take $-q_1,-q_2,\dots$ until it first drops below $s$. Then go back to the unused $p$'s, and so on. Each stage ends after finitely many terms by Step 1, and every stage uses at least one new term, so every term is eventually used and we have a genuine rearrangement. Step 4: after the first stage, the distance from the running sum to $s$ never exceeds the last term at which the direction switched, and these terms tend to $0$ by Step 2. So $t_k\to s$. For $+\infty$, use targets $1,2,3,\dots$, taking one negative term after each target is passed. For general $\alpha\lt\beta$, alternate between real targets $\alpha_k\to\alpha$ and $\beta_k\to\beta$. ∎
(a) Mertens' theorem. If at least one of the two series converges absolutely, then $\sum c_n$ converges and $\sum c_n=AB$.
(b) Cauchy. If both converge absolutely, so does $\sum c_n$.
(c) Abel (statement only; proof via Abel's theorem on power series, §5.2). If $\sum c_n$ converges at all, its sum is $AB$.
Without absolute convergence of one factor, $\sum c_n$ may diverge (Example 7(c)).
Proof. (a) Say $\sum|a_n|=\alpha\lt\infty$. Write $A_N,B_N,C_N$ for the partial sums up to index $N$. Collecting the products $a_kb_j$ with $k+j\le N$ by the index $k$ gives $$C_N=\sum_{k=0}^Na_kB_{N-k}=A_NB+\sum_{k=0}^Na_k\beta_{N-k},\qquad\beta_m=B_m-B\to0 .$$ It suffices to show $\gamma_N=\sum_{k=0}^Na_k\beta_{N-k}\to0$. Let $|\beta_m|\le M$ for all $m$ (a convergent sequence is bounded) and $\varepsilon\gt0$, and choose $N_0$ with $|\beta_m|\lt\varepsilon$ for $m\ge N_0$. For $N\gt N_0$, split according to whether $N-k\ge N_0$: $$|\gamma_N|\le\varepsilon\sum_{k=0}^{N-N_0}|a_k|+M\sum_{k=N-N_0+1}^{N}|a_k|\le\varepsilon\alpha+M\sum_{k\gt N-N_0}|a_k| .$$ As $N\to\infty$ the last sum is a tail of a convergent series and tends to $0$. So $\limsup|\gamma_N|\le\varepsilon\alpha$ for every $\varepsilon$, hence $\gamma_N\to0$ and $C_N\to AB$. (b) $\sum_{n\le N}|c_n|\le\sum_{k+j\le N}|a_k||b_j|$, and this is at most $\Big(\sum_{k\le N}|a_k|\Big)\Big(\sum_{j\le N}|b_j|\Big)$, because the triangle $k+j\le N$ lies inside the square $k,j\le N$. The partial sums are bounded, so $\sum|c_n|$ converges. ∎
- $a_n\to0$ is necessary, not sufficient: $\sum\frac1n$, $\sum\log\left(1+\frac1n\right)$ and $\sum\frac1{\sqrt n+\sqrt{n+1}}$ all diverge.
- An exponent above $1$ is not enough unless it stays away from $1$: $\sum\frac1{n^{1+1/n}}$ diverges (Example 2(e)).
- Limit comparison needs non-negative terms: $\frac{(-1)^n}{\sqrt n}$ and $\frac{(-1)^n}{\sqrt n}+\frac1n$ have ratio $\to1$, yet only the first gives a convergent series.
- Ratio and root tests are blind at $1$: $\sum\frac1n$ and $\sum\frac1{n^2}$ both give $1$ in both tests.
- $\limsup\frac{a_{n+1}}{a_n}\gt1$ does not imply divergence: the series with terms $2^{-n}$ (odd $n$) and $3^{-n}$ (even $n$) converges (Example 4(d)). Divergence follows only if $\frac{|a_{n+1}|}{|a_n|}\ge1$ eventually, or from $\limsup|a_n|^{1/n}\gt1$.
- Leibniz needs monotonicity: $\sum\frac{(-1)^n}{\sqrt n+(-1)^n}$ alternates with terms $\to0$, but diverges (Example 5(c)).
- Convergence of $\sum a_n$ says nothing about $\sum a_n^2$ or $\sum a_n^3$: $a_n=\frac{(-1)^n}{\sqrt n}$ defeats $\sum a_n^2$. For cubes, use the blocks $\left(x_k,-\frac{x_k}2,-\frac{x_k}2\right)$ with $x_k=k^{-1/3}$ (Exercise IV.2). For $a_n\ge0$, $\sum a_n^2$ does converge, but $\sum\sqrt{a_n}$ need not: $a_n=\frac1{n^2}$.
- $\sum a_n$ convergent and $b_n\to1$ do not make $\sum a_nb_n$ converge: Example 6(d). With absolute convergence and bounded $(b_n)$ it does.
- $na_n\to0$ needs monotonicity and is not sufficient: $a_n=\frac1n$ on squares and $\frac1{n^2}$ elsewhere gives a convergent series with $na_n\not\to0$, while $\frac1{n\log n}$ has $na_n\to0$ and a divergent series.
- Brackets can be inserted, not removed: $(1-1)+(1-1)+\cdots=0$, but $1-1+1-\cdots$ diverges.
- Order matters for conditionally convergent series: $1+\frac13-\frac12+\frac15+\frac17-\frac14+\cdots=\frac32\log2$, while the same terms in the original order give $\log2$. Even rearrangements that keep the positive terms and the negative terms in their own orders can change the sum.
- The Cauchy product of two convergent series can diverge: $\sum\frac{(-1)^n}{\sqrt{n+1}}$ with itself (Example 7(c)). One absolutely convergent factor is enough to prevent this (Mertens).
- A subseries of a convergent series need not converge: the even-indexed terms of $\sum\frac{(-1)^n}n$ form $\sum\frac1{2k}$. For non-negative or absolutely convergent series every subseries converges.
Exercises
- 1 Find $\displaystyle\sum_{n=1}^\infty\frac1{n(n+2)}$.
- 2 Find $\displaystyle\sum_{n=1}^\infty\frac{2n+1}{n^2(n+1)^2}$.
- 3 Find $\displaystyle\sum_{n=0}^\infty\frac{2^n+3^n}{6^n}$.
- 4 Find $\displaystyle\sum_{n=1}^\infty\frac{n^2}{2^n}$, using Example 1(b).
- 5 Find $\displaystyle\sum_{n=2}^\infty\log\Big(1-\frac1{n^2}\Big)$.
- 6 Show that $\displaystyle\sum\frac1{\sqrt n+\sqrt{n+1}}$ diverges by computing its partial sums.
- 7 Prove that $\displaystyle\sum_{n=1}^\infty\arctan\frac1{2n^2}=\frac\pi4$. (Show that $\arctan\frac1{2n^2}$ equals $\arctan(2n+1)-\arctan(2n-1)$.)
- 8 Let $a_n\ge0$ and let $\sum a_n$ converge. Prove that $\sum_ka_{n_k}$ converges for every subsequence $(a_{n_k})$. Show by an example that this fails without $a_n\ge0$.
- 1 $\displaystyle\sum\frac{n+1}{n^3+2}$
- 2 $\displaystyle\sum\frac1{\sqrt{n(n+1)}}$
- 3 $\displaystyle\sum\frac{\sqrt{n+1}-\sqrt n}{n}$
- 4 $\displaystyle\sum n^2e^{-\sqrt n}$
- 5 $\displaystyle\sum_{n\ge2}\frac1{(\log n)^{10}}$
- 6 $\displaystyle\sum_{n\ge3}\frac1{(\log n)^{\log\log n}}$
- 7 $\displaystyle\sum\frac{(n!)^2}{(2n)!}$
- 8 $\displaystyle\sum\frac{n^n}{3^n\,n!}$
- 9 $\displaystyle\sum\Big(1-\frac1n\Big)^{n^2}$
- 10 For which real $p$ does $\displaystyle\sum_{n\ge2}\frac1{n^p\log n}$ converge?
- 1 $\displaystyle\sum\frac{(-1)^nn}{n^2+1}$
- 2 $\displaystyle\sum(-1)^n\big(\sqrt{n+1}-\sqrt n\big)$
- 3 $\displaystyle\sum(-1)^n\sin\frac1n$
- 4 $\displaystyle\sum(-1)^n\Big(1-\cos\frac1n\Big)$
- 5 $\displaystyle\sum_{n\ge2}\frac{(-1)^n}{\log n}$
- 6 $\displaystyle\sum\frac{\cos(n\pi/3)}{n}$
- 7 Prove that $1-\frac12+\frac13-\cdots=\log2$, using $s_{2n}=H_{2n}-H_n$ and Theorem 2.27.
- 8 How many terms of $\sum\frac{(-1)^{n+1}}{n^3}$ guarantee an error below $10^{-4}$?
- 9 Find all real $x$ for which $\sum\frac{x^n}n$ converges, and say where the convergence is absolute.
- 10 Prove that $\displaystyle\sum_{n=1}^\infty\frac{(-1)^{\lfloor\sqrt n\rfloor}}n$ converges. (Group the terms with $k^2\le n\lt(k+1)^2$ and show that the block sums $b_k$ satisfy $\left|b_k-\frac2k\right|\le\frac3{k^2}$.)
- 1 Let $a_n\ge0$ and let $\sum a_n$ converge. Prove that $\sum a_n^2$, $\sum\frac{a_n}{1+a_n}$ and $\sum\sqrt{a_na_{n+1}}$ converge.
- 2 Give an example of a convergent series $\sum a_n$ for which $\sum a_n^3$ diverges.
- 3 Prove that if $\sum a_n$ converges absolutely and $(b_n)$ is bounded, then $\sum a_nb_n$ converges absolutely. Show that "absolutely" cannot be weakened to "conditionally".
- 4 Find the sum of $1+\frac13+\frac15-\frac12-\frac14$ $+\frac17+\frac19+\frac1{11}-\frac16-\frac18$ $+\cdots$ (three positive terms, then two negative).
- 5 Prove that the Cauchy product of $\sum_{n\ge0}\frac{x^n}{n!}$ and $\sum_{n\ge0}\frac{y^n}{n!}$ is $\sum_{n\ge0}\frac{(x+y)^n}{n!}$, and conclude that $E(x)E(y)=E(x+y)$ for $E(x)=\sum\frac{x^n}{n!}$.
- 6 Show that the Cauchy product of $\sum_{n\ge0}\frac{(-1)^n}{n+1}$ with itself has terms $c_n=(-1)^n\frac{2H_{n+1}}{n+2}$, that it converges, and find its sum.
- 7 Let $(a_n)$ be decreasing, $a_n\ge0$, and let $\sum a_n$ converge. Prove that $\sum n(a_n-a_{n+1})$ converges and has the same sum.
- 8 Describe a rearrangement of $1-\frac12+\frac13-\cdots$ that diverges to $+\infty$, and one whose partial sums have $\liminf=-1$ and $\limsup=2$.
- 9 Applied. A ball is dropped from a height of $2$ m. After each bounce it rises to $\frac34$ of its previous height. Find the total distance travelled and, with $g=9.81\ \text{m/s}^2$ and no air resistance, the total time until it comes to rest. (A fall from height $h$ takes $\sqrt{2h/g}$ seconds.)
E1. CSIR Part B GATE MA Which of the following series converges? (A) $\displaystyle\sum_{n\ge1}\Big(\frac1n-\log\Big(1+\frac1n\Big)\Big)$; (B) $\displaystyle\sum_{n\ge1}\big(e^{1/n}-1\big)$; (C) $\displaystyle\sum_{n\ge1}\frac1{\sqrt n+\log n}$; (D) $\displaystyle\sum_{n\ge2}\frac1{(\log n)^2}$.
E2. CSIR Part B The set of all $x\gt0$ for which $\displaystyle\sum_{n\ge1}\frac{n!\,x^n}{n^n}$ converges is (A) $(0,1)$; (B) $(0,e)$; (C) $(0,e]$; (D) $(0,\infty)$.
E3. CSIR Part B The terms of $1-\frac12+\frac13-\frac14+\cdots$ are rearranged as $1+\frac13+\frac15+\frac17-\frac12$ $+\frac19+\frac1{11}+\frac1{13}+\frac1{15}-\frac14$ $+\cdots$ (four positive terms, then one negative, each kind in its original order). The rearranged series (A) converges to $\log2$; (B) converges to $\frac32\log2$; (C) converges to $2\log2$; (D) diverges to $+\infty$.
E4. CSIR Part C Let $\sum a_n$ be a convergent series of real numbers. Which of the following are true? (One or more options may be correct.) (A) $\sum a_n^2$ converges. (B) $\sum\frac{a_n}n$ converges. (C) If $a_n\gt0$ for all $n$, then $\sum a_n^2$ converges. (D) $\sum(a_n+a_{n+1})$ converges.
E5. CSIR Part C GATE MA Let $a_n\gt0$ for all $n$. Which of the following are true? (A) If $\sum a_n$ converges, then $\sum\frac{\sqrt{a_n}}n$ converges. (B) If $\sum a_n$ converges, then $na_n\to0$. (C) If $\limsup\frac{a_{n+1}}{a_n}\gt1$, then $\sum a_n$ diverges. (D) If $\limsup a_n^{1/n}\lt1$, then $\sum a_n$ converges.
E6. CSIR Part C Let $\sum a_n$ be conditionally convergent. Which of the following are true? (A) The series of positive terms of $\sum a_n$ diverges. (B) Some rearrangement of $\sum a_n$ diverges to $+\infty$. (C) Every rearrangement that keeps the positive terms in their original order and the negative terms in their original order has the same sum as $\sum a_n$. (D) The Cauchy product of $\sum a_n$ with itself converges.
Suggested plan — 3 lectures × 60 min.
- Foundations and comparison. Open with the chapter teaser (harmonic vs $\sum\frac1{n^2}$). Definition 2.10, Theorem 2.21 (prove the Cauchy criterion and the $n$-th term test; discuss grouping with $1-1+1-\cdots$), geometric and telescoping series, Example 1 on the board. Theorem 2.23, the comparison tests, the condensation test with its picture of dyadic blocks, Theorem 2.26. Example 2(d), (e) and (f) make good "predict first" questions.
- Integral, root, ratio and Raabe. The integral test with the staircase picture and Euler's constant, Example 3(b). The root test with $\limsup$, the ratio test, and the comparison between them via Theorem 2.18. Project Interactive 1 and run Σ n!/nⁿ, then the odd/even series, then 1/n² against 1/n. Raabe's test with its two-line proof and Example 4(e). Finish with the strategy box.
- Signs, order and products. Absolute and conditional convergence, Leibniz (draw the nested partial sums), summation by parts, Dirichlet and Abel, Example 5(b) and (c). Riemann's theorem: let the class pick a target and project Interactive 2. Example 7(a) explains the predicted ratio $p/q$. Mertens' theorem and the divergent product of Example 7(c). Close with E4 or E6.
Discussion questions. (1) Every test for non-negative series ultimately proves that partial sums are bounded. Which comparison series sits behind each test? (2) Why can no test based on $a_{n+1}/a_n$ alone decide every series with ratio tending to $1$? (Compare $\frac1n$ with $\frac1{n\log^2n}$.) (3) Is there a "slowest divergent series", that is, one that every divergent positive series dominates? (No: if $\sum a_n$ diverges with partial sums $s_n$, then $\sum\frac{a_n}{s_n}$ also diverges.)
Misconceptions to watch for. Treating $a_n\to0$ as sufficient; using limit comparison with terms of changing sign; reading "$\limsup$ ratio $\gt1$" as divergence; applying Leibniz without checking monotonicity; believing that convergence of $\sum a_n$ implies convergence of $\sum a_n^2$; removing brackets freely; assuming the sum of a convergent series is independent of the order; multiplying two conditionally convergent series.
Suggested homework. I(1, 4, 5, 7, 8), II(2, 4, 5, 6, 8, 10), III(1, 2, 6, 8, 9, 10), IV(1, 2, 4, 6, 9); Exam corner E2, E4, E5.
Research question. The harmonic series diverges. What happens if we delete every term whose denominator contains the digit $9$? This gives the Kempner series $1+\frac12+\cdots+\frac18+\frac1{10}+\cdots+\frac1{18}+\frac1{20}+\cdots$. Does it converge, and if so, can its sum be computed?
Why it matters. The Kempner series converges, yet its partial sums creep upward so slowly that direct summation of billions of terms is far from the sum. It is a standard test case for convergence acceleration: methods that estimate the sum of a slowly convergent series from a few of its terms. The same ideas are used in computational physics (lattice sums in crystallography), in numerical number theory, and wherever an infinite sum must be evaluated to many digits with finite effort. It also shows that a divergent series can lose its divergence when a "thin-looking" set of terms is removed: almost all large integers contain a $9$.
What to do.
- Show that there are $8\cdot9^{k-1}$ integers with exactly $k$ digits and no digit $9$. Each is at least $10^{k-1}$, so their reciprocals sum to at most $8\left(\frac9{10}\right)^{k-1}$. Deduce convergence with sum at most $80$. This is a condensation-type argument in base $10$.
- Sharpen the bound by splitting by the leading digit $d\in\{1,\dots,8\}$: show that the sum is at most $10\left(1+\frac12+\cdots+\frac18\right)\approx27.18$.
- Compute the partial sums $K_m$ over all such $n\lt10^m$ for $m=1,\dots,7$ (or $8$ with vectorised code). Look at the differences $K_{m+1}-K_m$: roughly what ratio do they decrease by? Use this to extrapolate the sum with a geometric tail estimate.
- Compare your estimate with the known value $22.9206766\ldots$ How many terms would naive summation need to reach it to within $0.01$?
- Repeat steps 1–2 for "no digit $0$" or "no string $42$". Do the arguments still work?
Tools: Python (NumPy) or any language with fast loops; a spreadsheet is enough for steps 1–2.
Hand in: Your proofs of steps 1–2, a table of $K_m$ and $K_{m+1}-K_m$, your extrapolated sum with an honest error estimate, and two sentences explaining why "almost all integers contain a $9$" is consistent with convergence.
Going further. Read about Baillie's and Schmelzer–Baillie's algorithms, which compute Kempner-type sums to many digits by organising the terms by their digit blocks, and about Richardson extrapolation and the Euler–Maclaurin formula as general acceleration tools.
🧪 Self-check quiz
Q1. The series $\sum_{n\ge1}\frac1{n^p}$ converges exactly when
Q2. The series $\sum_{n\ge1}\frac{(-1)^{n+1}}{\sqrt n}$
Q3. For $\sum\frac{n^5}{3^n}$ the ratio $\frac{a_{n+1}}{a_n}$
Q4. Let $s=\sum_{n\ge1}\frac{(-1)^{n+1}}{n^2}$ and $s_{10}$ its tenth partial sum. The Leibniz estimate gives
Q5. If $\sum a_n$ converges conditionally, then by Riemann's theorem