Chapter 2 · Sequences and Series of Real Numbers

2.2 Subsequences, Bolzano–Weierstrass, lim sup and lim inf

Every bounded sequence, however wild, contains a convergent subsequence. Its largest and smallest subsequential limits — $\limsup$ and $\liminf$ — measure exactly how far it is from converging.

Most sequences do not converge. The terms of $\sin n$, of $(-1)^n\left(1+\tfrac1n\right)$ or of a population model in a chaotic regime wander for ever. Yet even such sequences have structure: inside them we can find subsequences that do converge, and the set of their limits — the places the sequence keeps returning to — tells us a great deal. The Bolzano–Weierstrass theorem guarantees that a bounded sequence always has at least one such place. This single fact is the engine behind the extreme value theorem, uniform continuity on closed bounded intervals, the Heine–Borel theorem and sequential compactness in Chapter 7.

The largest and smallest subsequential limits are the limit superior and limit inferior. Unlike the limit, they always exist (in $\overline{\mathbb{R}}$), they turn every "the limit exists" hypothesis into a usable inequality, and they are exactly what the root test, the Cauchy–Hadamard formula for the radius of convergence (§5.2) and many convergence proofs need. We finish with averages: Cesàro means, which can tame oscillating sequences, and the Stolz–Cesàro theorem, a discrete L'Hôpital rule. CSIR NET examines all of this heavily — computing $\limsup$ and $\liminf$, deciding which inequalities between them hold, and evaluating limits like $(n!)^{1/n}/n$ or $\tfrac1n\sum k^{1/k}$.

Subsequences and the Bolzano–Weierstrass Theorem

📘 Definition 2.6 — Subsequence
Let $(x_n)$ be a sequence and $n_1\lt n_2\lt n_3\lt\cdots$ a strictly increasing sequence of natural numbers. The sequence $(x_{n_k})_{k\ge1}=(x_{n_1},x_{n_2},x_{n_3},\dots)$ is a subsequence of $(x_n)$. By induction $n_k\ge k$ for every $k$. Examples: $(x_{2k})$, $(x_{2k-1})$, $(x_{k^2})$, $(x_{k!})$. Note that the order of the terms is kept and infinitely many terms are used; $(x_1,x_1,x_2,\dots)$ and $(x_2,x_1,x_3,x_4,\dots)$ are not subsequences.
📐 Theorem 2.11 — Subsequences of convergent sequences; divergence criteria
(a) If $x_n\to L$ (with $L\in\mathbb{R}$ or $L=\pm\infty$), then every subsequence $x_{n_k}\to L$.
(b) Divergence criteria. $(x_n)$ diverges if it has two subsequences with different limits, or a subsequence that is unbounded.
(c) $x_n\to L$ if and only if every subsequence of $(x_n)$ has a further subsequence converging to $L$.
(d) If $x_{2k}\to L$ and $x_{2k-1}\to L$, then $x_n\to L$. More generally, if finitely many subsequences together use every index and all converge to $L$, then $x_n\to L$.

Proof. (a) Let $\varepsilon\gt0$ and choose $N$ with $|x_n-L|\lt\varepsilon$ for $n\ge N$. For $k\ge N$ we have $n_k\ge k\ge N$, so $|x_{n_k}-L|\lt\varepsilon$. (The case $L=\pm\infty$ is identical with "$x_n\gt M$".)

(b) follows from (a) and uniqueness of limits: if $(x_n)$ converged to $L$, every subsequence would converge to the same $L$ and be bounded.

(c) "Only if" is (a). Conversely, suppose $x_n\not\to L$. Negating the definition, there is $\varepsilon_0\gt0$ such that for every $N$ some $n\ge N$ has $|x_n-L|\ge\varepsilon_0$. Choose $n_1$ with $|x_{n_1}-L|\ge\varepsilon_0$, then $n_2\gt n_1$ with the same property, and so on. Every term of the subsequence $(x_{n_k})$, hence of every further subsequence, is at distance $\ge\varepsilon_0$ from $L$, so no further subsequence converges to $L$.

(d) Given $\varepsilon\gt0$, choose $K$ with $|x_{2k}-L|\lt\varepsilon$ and $|x_{2k-1}-L|\lt\varepsilon$ for $k\ge K$. Then $|x_n-L|\lt\varepsilon$ for every $n\ge2K-1$. With finitely many subsequences, take the largest of the finitely many thresholds. ∎

The word "finitely" in (d) matters. Let $x_n=1$ if $n$ is prime and $x_n=0$ otherwise. For each prime $p$ the subsequence $(x_{pk})_{k\ge1}$ is $1,0,0,0,\dots$ and tends to $0$, and these infinitely many subsequences together use every index $n\ge2$ — yet $(x_n)$ diverges, because no single threshold works for all of them.

📐 Theorem 2.12 — Monotone subsequence theorem
Every sequence of real numbers has a monotone subsequence.

Proof. Call an index $m$ a peak if $x_m\ge x_n$ for every $n\gt m$ (the term $x_m$ is never exceeded later). Case 1: infinitely many peaks $m_1\lt m_2\lt\cdots$. Since each $m_k$ is a peak and $m_{k+1}\gt m_k$, we get $x_{m_1}\ge x_{m_2}\ge x_{m_3}\ge\cdots$: a decreasing subsequence. Case 2: finitely many peaks. Let $n_1$ be larger than every peak. Since $n_1$ is not a peak, some $n_2\gt n_1$ has $x_{n_2}\gt x_{n_1}$; since $n_2$ is not a peak either, some $n_3\gt n_2$ has $x_{n_3}\gt x_{n_2}$; continuing, we obtain a strictly increasing subsequence. ∎

📐 Theorem 2.13 — Bolzano–Weierstrass theorem
Every bounded sequence of real numbers has a convergent subsequence. If $a\le x_n\le b$ for all $n$, the limit of the subsequence lies in $[a,b]$.

Proof. By Theorem 2.12, $(x_n)$ has a monotone subsequence. It is bounded because $(x_n)$ is, so it converges by the monotone convergence theorem (§2.1). Its limit lies in $[a,b]$ by the order property of limits (§2.1, Theorem 2.2). ∎

Second proof (bisection — the one that generalises to $\mathbb{R}^n$). Put $I_1=[a,b]$; it contains $x_n$ for infinitely many $n$. Halve it: at least one closed half contains $x_n$ for infinitely many $n$; call it $I_2$. Continuing, we get nested closed intervals $I_k$ of length $(b-a)/2^{k-1}$, each containing infinitely many terms. Choose $n_1\lt n_2\lt\cdots$ with $x_{n_k}\in I_k$ (possible since infinitely many indices are available at each stage). By the nested interval property (§1.2) the $I_k$ meet in one point $x$, and $|x_{n_k}-x|\le(b-a)/2^{k-1}\to0$. In $\mathbb{R}^n$ cut a box into $2^n$ sub-boxes instead of two halves. ∎

📐 Theorem 2.14 — Consequences of Bolzano–Weierstrass
(a) If $(x_n)$ is bounded and every convergent subsequence of $(x_n)$ has the same limit $L$, then $x_n\to L$.
(b) A bounded sequence that does not converge has two subsequences converging to different limits.
(c) (Second proof of the Cauchy criterion.) Every Cauchy sequence of real numbers converges.

Proof. (a) Suppose $x_n\not\to L$. As in Theorem 2.11(c) there are $\varepsilon_0\gt0$ and a subsequence with $|x_{n_k}-L|\ge\varepsilon_0$ for all $k$. It is bounded, so by Bolzano–Weierstrass it has a further subsequence converging to some $L'$; this is also a convergent subsequence of $(x_n)$, and $|L'-L|\ge\varepsilon_0$ by the order property — contradicting the hypothesis. (b) By Bolzano–Weierstrass some subsequence converges, to $L$ say; by (a), if every convergent subsequence had limit $L$ the sequence would converge. (c) A Cauchy sequence is bounded (§2.1, Theorem 2.8), so some subsequence $x_{n_k}\to L$. Given $\varepsilon$, choose $N$ with $|x_m-x_n|\lt\varepsilon/2$ for $m,n\ge N$, and then $k$ with $n_k\ge N$ and $|x_{n_k}-L|\lt\varepsilon/2$. For $n\ge N$: $$|x_n-L|\le|x_n-x_{n_k}|+|x_{n_k}-L|\lt\varepsilon .\ ∎$$

Boundedness in (a) cannot be dropped: in $1,0,2,0,3,0,\dots$ every convergent subsequence tends to $0$, but the sequence is unbounded and diverges.

✏️ Example 1 — Subsequences detect divergence
(a) Show that $x_n=\cos\dfrac{n\pi}3$ diverges and find all its subsequential limits. (b) Show that $y_n=n^{(-1)^n}$ is neither convergent nor properly divergent. (c) Explain why every number in $[-1,1]$ is the limit of some subsequence of $(\sin n)$.
(a) Periodic The values repeat with period $6$: $\tfrac12,-\tfrac12,-1,-\tfrac12,\tfrac12,1,\dots$ So $x_{6k}=1\to1$ and $x_{6k+3}=-1\to-1$: two subsequences with different limits, hence divergence (Theorem 2.11(b)). Every subsequence takes infinitely often at least one of the four values $\pm\tfrac12,\pm1$, and a convergent subsequence is eventually constant here (consecutive distinct values differ by at least $\tfrac12$); the subsequential limits are exactly $\{-1,-\tfrac12,\tfrac12,1\}$.
(b) Two behaviours $y_{2k}=2k\to+\infty$ and $y_{2k-1}=\tfrac1{2k-1}\to0$. The first makes $(y_n)$ unbounded, so it does not converge; the second shows it does not tend to $+\infty$ (a properly divergent sequence has all subsequences tending to $+\infty$, Theorem 2.11(a)), nor to $-\infty$ (all terms are positive).
(c) Density Let $y=\sin\theta\in[-1,1]$. Because $\pi$ is irrational, the numbers $n-2\pi m$ ($n\in\mathbb{N}$, $m\in\mathbb{Z}$) come arbitrarily close to $\theta$ with $n$ as large as we like (the additive-subgroup argument of §1.2, Exercise III.7). Choose $n_1\lt n_2\lt\cdots$ and $m_k$ with $|n_k-2\pi m_k-\theta|\lt\tfrac1k$. Since $|\sin u-\sin v|\le|u-v|$, $$|\sin n_k-y|=|\sin(n_k-2\pi m_k)-\sin\theta|\lt\tfrac1k\to0.$$
Lesson A sequence can have a single, finitely many, or a whole interval of subsequential limits. The set of subsequential limits is the right "picture" of a divergent bounded sequence.
✏️ Example 2 — Prove or disprove
(a) An unbounded sequence has no convergent subsequence. (b) If every subsequence of $(x_n)$ has a convergent subsequence, then $(x_n)$ is bounded. (c) If $(x_{2n})$ and $(x_{3n})$ converge, then $(x_n)$ converges. (d) If $(x_{2n})$, $(x_{2n-1})$ and $(x_{3n})$ converge, then $(x_n)$ converges.
(a) False $1,0,2,0,3,0,\dots$ is unbounded, and the subsequence of zeros converges.
(b) True If $(x_n)$ were unbounded we could choose $n_1\lt n_2\lt\cdots$ with $|x_{n_k}|\gt k$; every subsequence of $(x_{n_k})$ is unbounded, so none converges.
(c) False Let $x_n=1$ if $n$ is a prime $\ge5$ and $x_n=0$ otherwise. Then $x_{2n}=0$ and $x_{3n}=0$ for $n\ge2$ (an even number $\gt2$ or a multiple of $3$ larger than $3$ is not prime), so both subsequences converge to $0$; but $x_p=1$ for the infinitely many primes $p$, so $(x_n)$ diverges. The two subsequences miss the indices $n\equiv\pm1\pmod6$.
(d) True $(x_{6n})$ is a subsequence of both $(x_{2n})$ and $(x_{3n})$, so their limits agree; $(x_{6n-3})$ is a subsequence of both $(x_{2n-1})$ and $(x_{3n})$, so those limits agree. Hence $\lim x_{2n}=\lim x_{2n-1}$, and $(x_n)$ converges by Theorem 2.11(d).

Subsequential Limits, lim sup and lim inf

📘 Definition 2.7 — Subsequential limit
A point $L\in\overline{\mathbb{R}}=\mathbb{R}\cup\{\pm\infty\}$ is a subsequential limit (or cluster value, or limit point of the sequence) of $(x_n)$ if some subsequence of $(x_n)$ converges to $L$. Equivalently, for real $L$: for every $\varepsilon\gt0$, $|x_n-L|\lt\varepsilon$ for infinitely many $n$. We write $\mathcal{S}(x_n)$ for the set of subsequential limits. Bolzano–Weierstrass says $\mathcal{S}(x_n)\cap\mathbb{R}\ne\varnothing$ for bounded sequences.

Do not confuse a subsequential limit with a limit point of the set $\{x_n\}$: the constant sequence $x_n=1$ has subsequential limit $1$, but the set $\{1\}$ has no limit points.

📘 Definition 2.8 — Limit superior and limit inferior

For a sequence $(x_n)$ put $u_n=\sup_{k\ge n}x_k$ and $\ell_n=\inf_{k\ge n}x_k$ (in $\overline{\mathbb{R}}$), the supremum and infimum of the $n$-th tail. As $n$ grows the tails shrink, so $(u_n)$ is decreasing and $(\ell_n)$ is increasing. Define

$$\limsup_{n\to\infty}x_n=\lim_{n\to\infty}u_n=\inf_{n\ge1}\,\sup_{k\ge n}x_k ,$$ $$\liminf_{n\to\infty}x_n=\lim_{n\to\infty}\ell_n=\sup_{n\ge1}\,\inf_{k\ge n}x_k .$$

These always exist in $\overline{\mathbb{R}}$. If $(x_n)$ is bounded they are real numbers; $\limsup x_n=+\infty$ iff $(x_n)$ is not bounded above, and $\limsup x_n=-\infty$ iff $x_n\to-\infty$. Other notations: $\overline{\lim}\,x_n$ and $\underline{\lim}\,x_n$.

📐 Theorem 2.15 — Three descriptions of lim sup
Let $(x_n)$ be bounded and $L\in\mathbb{R}$. The following are equivalent:
  1. $L=\limsup x_n$;
  2. for every $\varepsilon\gt0$: $x_n\lt L+\varepsilon$ for all but finitely many $n$, and $x_n\gt L-\varepsilon$ for infinitely many $n$;
  3. $L$ is the largest subsequential limit: $L\in\mathcal{S}(x_n)$ and $L'\le L$ for every $L'\in\mathcal{S}(x_n)$.
The mirror statements hold for $\liminf$ (the smallest subsequential limit: $x_n\gt L-\varepsilon$ eventually and $x_n\lt L+\varepsilon$ infinitely often). For unbounded sequences, $\limsup x_n=+\infty$ iff some subsequence tends to $+\infty$; so in all cases $\limsup x_n=\max\mathcal{S}(x_n)$ and $\liminf x_n=\min\mathcal{S}(x_n)$ in $\overline{\mathbb{R}}$.

Proof. (1)⇒(2). Let $L=\limsup x_n=\inf_nu_n$ and $\varepsilon\gt0$. Since $L+\varepsilon$ is not a lower bound of $\{u_n\}$, some $u_N\lt L+\varepsilon$, and then $x_n\le u_N\lt L+\varepsilon$ for all $n\ge N$. On the other hand $u_n\ge L\gt L-\varepsilon$ for every $n$, so the tail $\{x_k:k\ge n\}$ contains an element $\gt L-\varepsilon$; as $n$ is arbitrary, there are infinitely many such indices.

(2)⇒(3). Taking $\varepsilon=\tfrac1k$, all but finitely many $n$ satisfy $x_n\lt L+\tfrac1k$ and infinitely many satisfy $x_n\gt L-\tfrac1k$; so infinitely many $n$ satisfy both. Choose $n_1$ with $|x_{n_1}-L|\lt1$, then $n_2\gt n_1$ with $|x_{n_2}-L|\lt\tfrac12$, and so on: $x_{n_k}\to L$, so $L\in\mathcal{S}(x_n)$. If $x_{m_j}\to L'$, then since $x_{m_j}\lt L+\varepsilon$ for all large $j$, $L'\le L+\varepsilon$ for every $\varepsilon\gt0$, so $L'\le L$.

(3)⇒(1). By the two steps above, $L^\ast=\limsup x_n$ is itself the largest element of $\mathcal{S}(x_n)$. A set has at most one largest element, so $L=L^\ast$. For $\liminf$ apply everything to $(-x_n)$, using $\inf_{k\ge n}x_k=-\sup_{k\ge n}(-x_k)$. If $(x_n)$ is not bounded above, choose $n_1\lt n_2\lt\cdots$ with $x_{n_k}\gt k$: then $u_n=+\infty$ for all $n$ and a subsequence tends to $+\infty$; conversely a subsequence tending to $+\infty$ makes every $u_n=+\infty$. ∎

📐 Theorem 2.16 — Convergence via lim sup and lim inf
For every sequence, $\liminf x_n\le\limsup x_n$. A sequence converges to $L\in\mathbb{R}$ if and only if $\liminf x_n=\limsup x_n=L$; and $x_n\to+\infty$ iff $\liminf x_n=+\infty$. Moreover, for every $n$, $$\inf_k x_k\le\liminf x_n\le\limsup x_n\le\sup_kx_k .$$

Proof. $\ell_n\le u_n$ for every $n$, so the limits satisfy $\liminf\le\limsup$; also $\ell_1\le\ell_n$ and $u_n\le u_1$, which is the displayed chain. If $x_n\to L$, every subsequence tends to $L$, so $\mathcal{S}(x_n)=\{L\}$ and both are $L$ by Theorem 2.15. Conversely, if both equal $L\in\mathbb{R}$, then given $\varepsilon\gt0$ Theorem 2.15(2) gives $x_n\lt L+\varepsilon$ eventually (from $\limsup$) and $x_n\gt L-\varepsilon$ eventually (from $\liminf$), so $|x_n-L|\lt\varepsilon$ eventually. Finally $\liminf x_n=+\infty$ means $\ell_n=\inf_{k\ge n}x_k\to+\infty$, and $x_n\ge\ell_n$. ∎

🧭 Problem-solving strategy — computing $\limsup$ and $\liminf$
  1. Split the indices into finitely many classes on which the sequence behaves regularly — by parity, by $n\bmod m$ for periodic factors like $\cos\tfrac{2\pi n}{m}$, by "$n$ prime or not".
  2. Find the limit on each class. These are subsequential limits; since the classes cover all indices, they are all the subsequential limits (any subsequence has infinitely many terms in one class).
  3. Read off $\limsup=$ the largest and $\liminf=$ the smallest of them (Theorem 2.15).
  4. Separately compute $\sup$ and $\inf$ of the set of values if asked — they depend on early terms, $\limsup$ and $\liminf$ do not.
✏️ Example 3 — lim sup versus sup
For each sequence find $\limsup$, $\liminf$, $\sup$ and $\inf$: (a) $x_n=(-1)^n\left(1+\dfrac1n\right)$; (b) $y_n=n^{(-1)^n}$; (c) $z_n=\dfrac{n}{n+1}\cos\dfrac{2\pi n}{3}$.
(a) Even $n$: $1+\tfrac1n\downarrow1$; odd $n$: $-1-\tfrac1n\uparrow-1$. So $\mathcal{S}=\{-1,1\}$, $\limsup x_n=1$, $\liminf x_n=-1$. But $\sup x_n=x_2=\tfrac32$ and $\inf x_n=x_1=-2$: the extreme values are attained early and are not the $\limsup$/$\liminf$.
(b) $y_{2k}=2k\to\infty$, $y_{2k-1}=\tfrac1{2k-1}\to0$: $\limsup=+\infty$, $\liminf=0$, $\sup=+\infty$, $\inf=0$ (not attained, as all terms are positive).
(c) $\cos\tfrac{2\pi n}3=1$ if $3\mid n$ and $-\tfrac12$ otherwise. On $n=3k$: $z_n=\tfrac n{n+1}\uparrow1$; on the other indices: $z_n=-\tfrac12\cdot\tfrac n{n+1}\downarrow-\tfrac12$. So $\limsup=1$, $\liminf=-\tfrac12$. Here $\sup z_n=1$ and $\inf z_n=-\tfrac12$ as well, neither attained (the terms approach them strictly from inside).
Check Theorem 2.16's chain $\inf\le\liminf\le\limsup\le\sup$: (a) $-2\le-1\le1\le\tfrac32$ ✓.
Interactive 1 — Tails Shrinking onto lim sup and lim inf

The dots are $x_1,\dots,x_{60}$. Grey dots are discarded; the tail $\{x_k:k\ge m\}$ is in colour. The red line is $u_m=\sup_{k\ge m}x_k$ and the blue line is $\ell_m=\inf_{k\ge m}x_k$ (computed over $20\,000$ terms). As $m$ grows, $u_m$ can only fall and $\ell_m$ can only rise; they settle on $\limsup$ and $\liminf$ (dashed purple) — which meet exactly when the sequence converges. Press Play, or drag the slider.

m—
uₘ = sup of tail—
ℓₘ = inf of tail—
lim sup / lim inf—
subsequential limits—

Arithmetic of lim sup and lim inf; Ratios and Roots

📐 Theorem 2.17 — Properties of lim sup and lim inf
Let $(x_n)$, $(y_n)$ be bounded sequences. Then
  1. $\limsup(-x_n)=-\liminf x_n$; and for $c\gt0$, $\limsup(cx_n)=c\limsup x_n$;
  2. if $x_n\le y_n$ eventually, then $\limsup x_n\le\limsup y_n$ and $\liminf x_n\le\liminf y_n$;
  3. if $y_n\to y$, then $\limsup(x_n+y_n)=\limsup x_n+y$;
  4. the chain of inequalities $$\liminf x_n+\liminf y_n\le\liminf(x_n+y_n),$$ $$\liminf(x_n+y_n)\le\liminf x_n+\limsup y_n\le\limsup(x_n+y_n),$$ $$\limsup(x_n+y_n)\le\limsup x_n+\limsup y_n ;$$
  5. if $x_n,y_n\ge0$, then $\limsup(x_ny_n)\le\limsup x_n\cdot\limsup y_n$; and if $y_n\to y\ge0$, then $\limsup(x_ny_n)=y\limsup x_n$.
All inequalities in (4) and (5) can be strict.

Proof. Write $u_n(x)=\sup_{k\ge n}x_k$. (1) $\sup_{k\ge n}(-x_k)=-\inf_{k\ge n}x_k$ and $\sup_{k\ge n}cx_k=c\,u_n(x)$ for $c\gt0$; let $n\to\infty$. (2) For $n$ beyond the point where $x_k\le y_k$, $u_n(x)\le u_n(y)$; let $n\to\infty$. The $\liminf$ case is the same with infima.

(3) Given $\varepsilon\gt0$, eventually $y-\varepsilon\lt y_k\lt y+\varepsilon$, so for large $n$ $$u_n(x)+y-\varepsilon\le\sup_{k\ge n}(x_k+y_k)\le u_n(x)+y+\varepsilon .$$ Letting $n\to\infty$ and then $\varepsilon\to0$ gives the claim.

(4) Last inequality: for $k\ge n$, $x_k+y_k\le u_n(x)+u_n(y)$, so $u_n(x+y)\le u_n(x)+u_n(y)$; let $n\to\infty$. Third: choose a subsequence with $y_{n_k}\to\limsup y_n$ (Theorem 2.15). By (3) applied to the subsequences, and since the $\limsup$ of a subsequence is at most that of the sequence, $$\limsup(x_n+y_n)\ge\limsup_k(x_{n_k}+y_{n_k})=\limsup_kx_{n_k}+\limsup y_n .$$ Finally $\limsup_kx_{n_k}\ge\liminf_kx_{n_k}\ge\liminf x_n$ (every subsequential limit of the subsequence is one of the sequence). First and second: apply the last and the third inequality to $(-x_n),(-y_n)$ and use (1).

(5) For $k\ge n$, $0\le x_ky_k\le u_n(x)u_n(y)$, so $u_n(xy)\le u_n(x)u_n(y)$, and the right side tends to $\limsup x_n\limsup y_n$ (product of two convergent sequences). If $y_n\to y$ and $|x_n|\le M$, then $|x_ny_n-x_ny|\le M|y_n-y|\to0$, so by (3) $\limsup(x_ny_n)=\limsup(yx_n)$, which is $y\limsup x_n$ by (1) (trivially if $y=0$). ∎

✏️ Example 4 — When the inequalities are strict
(a) With $x_n=(-1)^n$ and $y_n=(-1)^{n+1}$, evaluate every term of the chain in Theorem 2.17(4). (b) Find bounded non-negative sequences with $\limsup(x_ny_n)\lt\limsup x_n\cdot\limsup y_n$. (c) Show that Theorem 2.17(5) fails without non-negativity.
(a) $\liminf x=\liminf y=-1$, $\limsup x=\limsup y=1$, and $x_n+y_n=0$. The chain reads $$-2\lt0=0=0\lt2 :$$ the outer inequalities are strict, the middle ones are equalities here.
(b) $x_n=2,1,2,1,\dots$ and $y_n=1,2,1,2,\dots$: $x_ny_n=2$ for all $n$, so $\limsup(x_ny_n)=2\lt2\cdot2=4$. The two $\limsup$'s are attained along different subsequences.
(c) $x_n=y_n=-2,0,-2,0,\dots$: $\limsup x_n=\limsup y_n=0$, but $x_ny_n=4,0,4,0,\dots$ has $\limsup=4\gt0\cdot0$.
Rule of thumb $\limsup$ is subadditive and $\liminf$ superadditive; equality holds as soon as one of the two sequences converges.
📐 Theorem 2.18 — Ratios control roots
For every sequence of positive numbers $(x_n)$, $$\liminf\frac{x_{n+1}}{x_n}\le\liminf x_n^{1/n}\le\limsup x_n^{1/n}\le\limsup\frac{x_{n+1}}{x_n}$$ (in $\overline{\mathbb{R}}$). In particular, if $\dfrac{x_{n+1}}{x_n}\to L\in[0,\infty]$, then $x_n^{1/n}\to L$.

Proof. The middle inequality is Theorem 2.16. For the right one, let $\beta=\limsup x_{n+1}/x_n$; if $\beta=+\infty$ there is nothing to prove. Take any $\beta'\gt\beta$. By Theorem 2.15(2) there is $N$ with $x_{n+1}/x_n\lt\beta'$ for all $n\ge N$, and by induction $$x_n\le x_N\beta'^{\,n-N}=C\beta'^{\,n}\quad(n\ge N),\qquad C=x_N\beta'^{-N}\gt0 .$$ Hence $x_n^{1/n}\le C^{1/n}\beta'$, and $C^{1/n}\to1$ (§2.1, Theorem 2.5), so $\limsup x_n^{1/n}\le\lim C^{1/n}\beta'=\beta'$ by Theorem 2.17(2). As $\beta'\gt\beta$ was arbitrary, $\limsup x_n^{1/n}\le\beta$. The left inequality is proved the same way with $\alpha'\lt\alpha=\liminf x_{n+1}/x_n$ and $x_n\ge C\alpha'^{\,n}$ (nothing to prove if $\alpha=0$). If the ratios converge to $L$, the outer terms are both $L$, so the root converges to $L$ by Theorem 2.16. ∎

✏️ Example 5 — Roots through ratios
Find (a) $\lim\dfrac{(n!)^{1/n}}{n}$; (b) $\lim\dbinom{2n}{n}^{1/n}$. (c) For $x_n=2^{-n+(-1)^n}$, compute both sides of Theorem 2.18 and show that the root test can succeed where the ratio test fails.
(a) Rewrite as a root $\dfrac{(n!)^{1/n}}{n}=a_n^{1/n}$ with $a_n=\dfrac{n!}{n^n}$. Then $$\frac{a_{n+1}}{a_n}=\frac{(n+1)!\,n^n}{(n+1)^{n+1}n!}=\left(\frac{n}{n+1}\right)^n=\frac1{(1+1/n)^n}\to\frac1e .$$ By Theorem 2.18 the limit is $\dfrac1e$. (So $n!$ grows like $(n/e)^n$ up to factors whose $n$-th root tends to $1$ — a first glimpse of Stirling's formula. For $n=2000$ the quotient is $0.3687$; $1/e=0.3679$.)
(b) $\dfrac{\binom{2n+2}{n+1}}{\binom{2n}{n}}=\dfrac{(2n+2)(2n+1)}{(n+1)^2}=\dfrac{2(2n+1)}{n+1}\to4$, so $\binom{2n}n^{1/n}\to4$.
(c) Ratios oscillate $\dfrac{x_{n+1}}{x_n}=2^{-1+(-1)^{n+1}-(-1)^n}$, which is $2^{-3}=\tfrac18$ for even $n$ and $2^{1}=2$ for odd $n$. So $\liminf=\tfrac18$ and $\limsup=2$: the ratio test (§2.3) is inconclusive.
Roots converge $x_n^{1/n}=2^{-1+(-1)^n/n}\to\tfrac12$. The chain reads $\tfrac18\le\tfrac12=\tfrac12\le2$ ✓, and the root test will show $\sum x_n$ converges.

Cesàro Means and the Stolz–Cesàro Theorem

📘 Definition 2.9 — Cesàro means
The Cesàro means (arithmetic means) of $(x_n)$ are $$\sigma_n=\frac{x_1+x_2+\cdots+x_n}{n}.$$ If $\sigma_n\to L$ we say $(x_n)$ is Cesàro convergent to $L$ (for a series, applied to its partial sums: Cesàro summable).
📐 Theorem 2.19 — Cesàro's theorem (arithmetic and geometric means)
(a) For every real sequence, $$\liminf x_n\le\liminf\sigma_n\le\limsup\sigma_n\le\limsup x_n .$$ In particular, if $x_n\to L$ ($L\in\mathbb{R}$ or $\pm\infty$), then $\sigma_n\to L$.
(b) If $x_n\gt0$ and $x_n\to L\in[0,\infty)$, then the geometric means $g_n=(x_1x_2\cdots x_n)^{1/n}\to L$.
The converses fail: $(-1)^n$ is Cesàro convergent to $0$ but divergent.

Proof. (a) We prove the last inequality; the first follows by applying it to $(-x_n)$. Let $\beta=\limsup x_n$; assume $\beta\lt+\infty$ and take any real $\beta'\gt\beta$. There is $N$ with $x_k\lt\beta'$ for $k\gt N$. For $n\gt N$, $$\sigma_n=\frac{x_1+\cdots+x_N}{n}+\frac{x_{N+1}+\cdots+x_n}{n}\lt\frac{C_N}{n}+\frac{n-N}{n}\beta' ,$$ where $C_N=x_1+\cdots+x_N$ is fixed. The right side tends to $\beta'$, so $\limsup\sigma_n\le\beta'$ (Theorem 2.17(2)); since $\beta'\gt\beta$ is arbitrary, $\limsup\sigma_n\le\beta$. If $x_n\to L$, the two outer terms equal $L$, which forces $\sigma_n\to L$ (Theorem 2.16).

(b) Recall the inequalities between the harmonic, geometric and arithmetic means of positive numbers: $$\frac{n}{1/x_1+\cdots+1/x_n}\le g_n\le\sigma_n .$$ If $L=0$, then $0\lt g_n\le\sigma_n\to0$. If $L\gt0$, then $1/x_n\to1/L$, so by (a) the Cesàro means of $(1/x_n)$ tend to $1/L$, and the harmonic mean on the left tends to $L$; so does $\sigma_n$; squeeze. ∎

Interactive 2 — Averaging Tames (Some) Oscillation

Faint dots: the terms $x_n$. Red curve: the Cesàro means $\sigma_n$. The horizontal axis is logarithmic ($n$ from $1$ to about $4000$), so slow behaviour is visible. Theorem 2.19 traps $\liminf\sigma_n$ and $\limsup\sigma_n$ between $\liminf x_n$ and $\limsup x_n$ (the dashed lines). Some divergent sequences have convergent means — but not all.

n—
xₙ—
σₙ—
lim inf xₙ, lim sup xₙ—
behaviour of σₙ—
📐 Theorem 2.20 — Stolz–Cesàro theorem
(a) ($\tfrac{\cdot}{\infty}$ form) Let $(b_n)$ be strictly increasing with $b_n\to+\infty$, and suppose $$\frac{a_{n+1}-a_n}{b_{n+1}-b_n}\to L\qquad(L\in\mathbb{R}\text{ or }L=\pm\infty).$$ Then $\dfrac{a_n}{b_n}\to L$.
(b) ($\tfrac00$ form) If $a_n\to0$, $b_n\to0$, $(b_n)$ is strictly decreasing and $\dfrac{a_{n+1}-a_n}{b_{n+1}-b_n}\to L\in\mathbb{R}$, then $\dfrac{a_n}{b_n}\to L$.
The converse is false. Cesàro's theorem is the case $b_n=n$, $a_n=x_1+\cdots+x_n$.

Proof of (a) for real $L$. Let $\varepsilon\gt0$ and choose $N$ with $b_N\gt0$ and $$L-\varepsilon\lt\frac{a_{k+1}-a_k}{b_{k+1}-b_k}\lt L+\varepsilon\qquad(k\ge N).$$ Multiply by $b_{k+1}-b_k\gt0$ and add these inequalities for $k=N,\dots,n-1$; the sums telescope: $$(L-\varepsilon)(b_n-b_N)\lt a_n-a_N\lt(L+\varepsilon)(b_n-b_N).$$ Divide by $b_n\gt0$: $$(L-\varepsilon)\Big(1-\frac{b_N}{b_n}\Big)+\frac{a_N}{b_n}\lt\frac{a_n}{b_n}\lt(L+\varepsilon)\Big(1-\frac{b_N}{b_n}\Big)+\frac{a_N}{b_n}.$$ With $N$ fixed and $n\to\infty$, $b_N/b_n\to0$ and $a_N/b_n\to0$. Hence $L-\varepsilon\le\liminf\frac{a_n}{b_n}\le\limsup\frac{a_n}{b_n}\le L+\varepsilon$, for every $\varepsilon\gt0$, and $a_n/b_n\to L$. For $L=+\infty$ replace $L\pm\varepsilon$ by a large $M$ and keep only the left inequality. Part (b) is proved the same way, summing from $k=n$ to $m-1$ and letting $m\to\infty$ first. ∎

The converse fails: $a_n=(-1)^n$, $b_n=n$ give $a_n/b_n\to0$, while $(a_{n+1}-a_n)/(b_{n+1}-b_n)=\pm2$ diverges.

✏️ Example 6 — Stolz–Cesàro at work
Find (a) $\lim\dfrac{1^k+2^k+\cdots+n^k}{n^{k+1}}$ for fixed $k\in\mathbb{N}$; (b) $\lim\dfrac{1+\frac12+\cdots+\frac1n}{\log n}$; (c) for $x_1=1$, $x_{n+1}=\sin x_n$, the limit of $\sqrt n\,x_n$, given that $\dfrac1{\sin^2x}-\dfrac1{x^2}\to\dfrac13$ as $x\to0$ (Taylor's theorem, §3.3).
(a) Differences $a_n=\sum_{j\le n}j^k$, $b_n=n^{k+1}$ (strictly increasing to $\infty$). $$\frac{a_{n+1}-a_n}{b_{n+1}-b_n}=\frac{(n+1)^k}{(n+1)^{k+1}-n^{k+1}} .$$ By the binomial theorem the denominator is $(k+1)n^k+(\text{terms of degree}\lt k)$ and the numerator is $n^k+\cdots$, so the quotient tends to $\dfrac1{k+1}$. Hence the limit is $\dfrac1{k+1}$ (for $k=3$: $\tfrac14$) — the discrete version of $\int_0^1x^k\,dx$.
(b) $a_n=H_n$, $b_n=\log n$ (strictly increasing to $\infty$). $$\frac{H_{n+1}-H_n}{\log(n+1)-\log n}=\frac{1}{(n+1)\log(1+\frac1n)}=\frac{1}{\log\left(1+\frac1n\right)^{n+1}}\to\frac1{\log e}=1,$$ using $(1+\tfrac1n)^{n+1}\to e$ (§2.1) and continuity of $\log$ (§3.1). So $H_n\sim\log n$.
(c) Linearise the recursion $0\lt\sin x\lt x$ on $(0,1]$, so $(x_n)$ decreases and is positive; its limit $L$ satisfies $L=\sin L$, so $L=0$. Put $t_n=1/x_n^2$. Then $t_{n+1}-t_n=\dfrac1{\sin^2x_n}-\dfrac1{x_n^2}\to\dfrac13$. By Stolz–Cesàro with $b_n=n$: $\dfrac{t_n}{n}\to\dfrac13$, i.e. $nx_n^2\to3$.
Answer and check $\sqrt n\,x_n\to\sqrt3\approx1.7321$ (square root is continuous). Numerically, after $10^6$ iterations $\sqrt n\,x_n=1.73204$. This "$x_n\sim\sqrt{3/n}$" is a favourite Part C fact.
⚠️ Counterexamples and pitfalls
  1. $\limsup\ne\sup$: $(-1)^n(1+\tfrac1n)$ has $\sup=\tfrac32$ but $\limsup=1$. Changing finitely many terms changes $\sup$, never $\limsup$.
  2. A convergent subsequence does not make the sequence converge: $(-1)^n$ has the convergent subsequence $(x_{2n})$.
  3. Two subsequences covering "most" indices are not enough: $(x_{2n})$ and $(x_{3n})$ can converge while $(x_n)$ diverges (Example 2(c)); $(x_{2n})$, $(x_{2n-1})$ together suffice.
  4. Unbounded sequences may have convergent subsequences: $1,0,2,0,3,0,\dots$; and a unique finite subsequential limit gives convergence only for bounded sequences.
  5. Bolzano–Weierstrass needs completeness: in $\mathbb{Q}$ the bounded sequence of decimal truncations of $\sqrt2$ has no subsequence converging in $\mathbb{Q}$; in $(0,1)$ the sequence $\tfrac1{n+1}$ has no subsequence converging in $(0,1)$.
  6. $\limsup$ is only subadditive: $x_n=(-1)^n$, $y_n=-x_n$ give $\limsup(x_n+y_n)=0\lt2$.
  7. Products need signs: $\limsup(x_ny_n)\le\limsup x_n\limsup y_n$ holds for $x_n,y_n\ge0$; it fails for $x_n=y_n=-2,0,-2,0,\dots$
  8. The set of subsequential limits is closed: it cannot be $\{1/n:n\in\mathbb{N}\}$ without also containing $0$; it can be $\{0\}\cup\{1/n\}$, or all of $[0,1]$ (an enumeration of the rationals in $[0,1]$).
  9. Cesàro convergence is weaker than convergence: $(-1)^n$; and some bounded sequences are not even Cesàro convergent (the block sequence of Interactive 2: $\liminf\sigma_n=\tfrac13$, $\limsup\sigma_n=\tfrac23$).
  10. Roots can converge when ratios do not: $x_n=2^{-n+(-1)^n}$ (Example 5(c)); the reverse never happens (Theorem 2.18).
  11. Stolz–Cesàro has no converse: $a_n=(-1)^n$, $b_n=n$. And it needs $b_n$ strictly monotone and unbounded in the $\tfrac{\cdot}{\infty}$ form.
  12. Subsequential limit of a sequence $\ne$ limit point of its set of values: the constant sequence $1,1,1,\dots$ has subsequential limit $1$, but $\{1\}$ has no limit point.

Exercises

I
Subsequences and Bolzano–Weierstrass.
  • 1 Use subsequences to prove that $\cos\dfrac{n\pi}{3}+\dfrac1n$ and $n\sin\dfrac{n\pi}2$ diverge.
  • 2 Prove: $x_n\to L$ if and only if $x_{2n}\to L$ and $x_{2n-1}\to L$.
  • 3 Give a sequence whose set of subsequential limits is exactly $\{0,1,2\}$.
  • 4 Give a sequence whose set of real subsequential limits is $\mathbb{N}$, and one whose set of subsequential limits is $[0,1]$.
  • 5 Prove that the set of real subsequential limits of any sequence is closed: if each $L_j$ is a subsequential limit and $L_j\to L$, then $L$ is a subsequential limit. Can this set be $\{1/n:n\in\mathbb{N}\}$?
  • 6 Prove that an unbounded sequence has a subsequence tending to $+\infty$ or to $-\infty$.
  • 7 Find a monotone subsequence of $x_n=\dfrac{(-1)^n n}{n+1}$, and identify the peaks (Theorem 2.12) of this sequence.
  • 8 Let $(x_n)$ be bounded and suppose $x_{n+1}-x_n\to0$. Prove that the set of subsequential limits is a closed interval (possibly a single point). Give an example where it is $[-1,1]$.
1 $x_{6k}=1+\tfrac1{6k}\to1$ and $x_{6k+3}=-1+\tfrac1{6k+3}\to-1$. For $y_n=n\sin\tfrac{n\pi}2$: $y_{4k+1}=4k+1\to\infty$ (unbounded subsequence) and $y_{2k}=0$.
2 "Only if": Theorem 2.11(a). "If": given $\varepsilon$, choose $K$ working for both subsequences; then $|x_n-L|\lt\varepsilon$ for $n\ge2K$ (Theorem 2.11(d)).
3 $x_n=n\bmod3$: $1,2,0,1,2,0,\dots$ Each residue class gives a constant subsequence, and any subsequence has infinitely many terms in one class, so $\mathcal{S}=\{0,1,2\}$.
4 $1;\ 1,2;\ 1,2,3;\ 1,2,3,4;\dots$ takes every value $m$ infinitely often, so every $m\in\mathbb{N}$ is a subsequential limit; a convergent subsequence of integers is eventually constant, so there are no others in $\mathbb{R}$ ($+\infty$ is one in $\overline{\mathbb{R}}$). For $[0,1]$: enumerate $\mathbb{Q}\cap[0,1]$ as $(r_n)$; every $y\in[0,1]$ has rationals within $\tfrac1k$ of it with arbitrarily large index (each interval contains infinitely many rationals), so $y\in\mathcal{S}$; and no limit lies outside $[0,1]$.
5 Choose $n_1$ with $|x_{n_1}-L_1|\lt1$ (possible, $L_1$ is a subsequential limit); then $n_2\gt n_1$ with $|x_{n_2}-L_2|\lt\tfrac12$, etc. Then $|x_{n_j}-L|\le\tfrac1j+|L_j-L|\to0$. So $\{1/n\}$ is impossible: $0$ would have to belong to the set.
6 If not bounded above, choose $n_1\lt n_2\lt\cdots$ with $x_{n_k}\gt k$ (each tail is also unbounded above); then $x_{n_k}\to+\infty$. Otherwise it is unbounded below, and symmetrically some subsequence tends to $-\infty$.
7 Even terms $\tfrac{2k}{2k+1}$ increase to $1$: a monotone subsequence. Peaks: every tail has supremum $1$, not attained, so a peak $m$ would need $x_m\ge1$ — impossible. There are no peaks, and the proof of Theorem 2.12 produces an increasing subsequence, as found.
8 Let $a=\liminf x_n$, $b=\limsup x_n$ (finite, as the sequence is bounded); if $a=b$ the sequence converges. Let $a\lt c\lt b$ and $\varepsilon\gt0$, and choose $N$ with $|x_{n+1}-x_n|\lt\varepsilon$ for $n\ge N$. By Theorem 2.15, infinitely many $n\ge N$ have $x_n\lt c$ and infinitely many have $x_n\gt c$. After each index with $x_n\lt c$ there is a first later index $j$ with $x_j\ge c$; then $x_{j-1}\lt c\le x_j\lt x_{j-1}+\varepsilon$, so $|x_j-c|\lt\varepsilon$. This happens for infinitely many $j$, so $c$ is a subsequential limit. Together with $a,b\in\mathcal{S}$ and $\mathcal{S}\subseteq[a,b]$, $\mathcal{S}=[a,b]$. Example: $x_n=\sin\sqrt n$ (steps $\le\sqrt{n+1}-\sqrt n\to0$, and $\sqrt n$ passes every $\tfrac\pi2+2\pi m$ and $-\tfrac\pi2+2\pi m$ closely).
II
Computing lim sup and lim inf. In 1–6 find $\limsup$, $\liminf$, $\sup$ and $\inf$.
  • 1 $x_n=(-1)^n+\dfrac1n$
  • 2 $x_n=(-1)^n\left(1+\dfrac1n\right)^n$
  • 3 $x_n=\sin\dfrac{n\pi}2+\cos n\pi$
  • 4 $x_n=n\sin\dfrac{n\pi}{2}$
  • 5 $x_n=\dfrac{n\bmod5}{5}+\dfrac{(-1)^n}n$
  • 6 $x_n=\left(1+\dfrac{(-1)^n}{n}\right)^n$
  • 7 Prove $\limsup(-x_n)=-\liminf x_n$ directly from Definition 2.8.
  • 8 Prove: if $\limsup x_n\lt1$ then $x_n\lt1$ for all large $n$; and if $x_n\lt1$ for all $n$, then only $\limsup x_n\le1$ follows.
  • 9 Prove that $\limsup\sin n=1$ and $\liminf\sin n=-1$, while $\sin n\ne\pm1$ for every $n\in\mathbb{N}$.
1 Even: $1+\tfrac1n\downarrow1$; odd: $-1+\tfrac1n\downarrow-1$ (from $0$ at $n=1$). $\limsup=1$, $\liminf=-1$, $\sup=x_2=\tfrac32$, $\inf=-1$ (not attained).
2 $e_n=(1+\tfrac1n)^n\uparrow e$. Even: $e_n\to e$; odd: $-e_n\to-e$. $\limsup=e$, $\liminf=-e$; $\sup=e$ and $\inf=-e$, neither attained.
3 Period $4$: $n\equiv1$: $1-1=0$; $n\equiv2$: $0+1=1$; $n\equiv3$: $-1-1=-2$; $n\equiv0$: $0+1=1$. $\limsup=\sup=1$, $\liminf=\inf=-2$ (all attained).
4 $x_{4k+1}=4k+1\to\infty$, $x_{4k+3}=-(4k+3)\to-\infty$, $x_{2k}=0$. $\limsup=\sup=+\infty$, $\liminf=\inf=-\infty$; $0$ is also a subsequential limit.
5 On each residue class $r$ the terms tend to $\tfrac r5$: $\mathcal{S}=\{0,\tfrac15,\tfrac25,\tfrac35,\tfrac45\}$, $\limsup=\tfrac45$, $\liminf=0$. $\sup=x_4=\tfrac45+\tfrac14=1.05$ and $\inf=x_1=\tfrac15-1=-0.8$, both attained.
6 Even $n$: $(1+\tfrac1n)^n\to e$ (increasing). Odd $n$: $(1-\tfrac1n)^n\to\tfrac1e$ (§2.1, Exercise II.8). $\limsup=e$, $\liminf=\tfrac1e$; $\sup=e$ (not attained), $\inf=x_1=0$.
7 $\sup_{k\ge n}(-x_k)=-\inf_{k\ge n}x_k$ for each $n$ (Theorem 1.21 of §1.2 applied to the tail). Take limits as $n\to\infty$ (a decreasing sequence on the left, the negative of an increasing one on the right).
8 Take $\varepsilon=1-\limsup x_n\gt0$ in Theorem 2.15(2): eventually $x_n\lt\limsup x_n+\varepsilon=1$. For the second part, $x_n=1-\tfrac1n\lt1$ has $\limsup=1$; Theorem 2.17(2) gives $\le1$.
9 By Example 1(c) every $y\in[-1,1]$ is a subsequential limit, in particular $\pm1$; and all values lie in $[-1,1]$, so $\limsup=1$, $\liminf=-1$. If $\sin n=1$ then $n=\tfrac\pi2+2\pi m$, so $\pi=\tfrac{2n}{1+4m}$ would be rational — impossible; similarly for $-1$.
III
Inequalities, ratios and roots.
  • 1 Prove $\limsup(x_n+y_n)\le\limsup x_n+\limsup y_n$ for bounded sequences, and give an example with strict inequality in which neither sequence is periodic.
  • 2 Prove: if $y_n\to y$ then $\liminf(x_n+y_n)=\liminf x_n+y$.
  • 3 Prove: if $x_n\to x\gt0$ and $(y_n)$ is bounded, then $\limsup(x_ny_n)=x\limsup y_n$. What if $x\lt0$?
  • 4 Find $\lim\left(\dfrac{n^n}{n!}\right)^{1/n}$.
  • 5 Find $\lim\left(\dfrac{(3n)!}{(n!)^3}\right)^{1/n}$.
  • 6 Give a positive sequence for which $x_n^{1/n}$ converges but $x_{n+1}/x_n$ does not, other than Example 5(c).
  • 7 Let $x_n\gt0$ and $\limsup x_n^{1/n}=\rho\lt1$. Prove $x_n\to0$. Is the same true if only $\liminf x_n^{1/n}\lt1$?
  • 8 Prove that $\limsup x_n=\lim_{n\to\infty}\sup_{k\ge n}x_k$ exists for every bounded sequence without using Theorem 2.15, and show that $\limsup x_n\le\limsup x_{n_k}$ is false in general for a subsequence while $\limsup x_{n_k}\le\limsup x_n$ always holds.
1 Proof as in Theorem 2.17(4): $\sup_{k\ge n}(x_k+y_k)\le u_n(x)+u_n(y)$. Example: $x_n=\sin n$, $y_n=-\sin n$: $\limsup(x_n+y_n)=0\lt1+1$.
2 Apply Theorem 2.17(3) to $(-x_n)$ and $(-y_n)$ and use $\liminf=-\limsup(-\,\cdot)$.
3 $|x_ny_n-xy_n|\le M|x_n-x|\to0$, so $\limsup(x_ny_n)=\limsup(xy_n)=x\limsup y_n$ (Theorem 2.17(3), (1)). If $x\lt0$: $\limsup(xy_n)=x\liminf y_n$.
4 $a_n=\tfrac{n^n}{n!}$: $\tfrac{a_{n+1}}{a_n}=\left(1+\tfrac1n\right)^n\to e$, so the limit is $e$ (Theorem 2.18) — consistent with Example 5(a).
5 $\dfrac{a_{n+1}}{a_n}=\dfrac{(3n+3)(3n+2)(3n+1)}{(n+1)^3}\to27$: the limit is $27$.
6 $x_n=2+(-1)^n$: ratios alternate between $3$ and $\tfrac13$, while $1\le x_n^{1/n}\le3^{1/n}\to1$.
7 Choose $\rho\lt r\lt1$; eventually $x_n^{1/n}\lt r$ (Theorem 2.15(2)), so $0\lt x_n\lt r^n\to0$. No: $x_n=1$ for even $n$ and $2^{-n}$ for odd $n$ has $\liminf x_n^{1/n}=\tfrac12$ but does not tend to $0$.
8 $u_n=\sup_{k\ge n}x_k$ is decreasing (sup over a smaller set) and bounded below by $\inf x_k$, so it converges (monotone convergence theorem). For a subsequence, $\sup_{j\ge k}x_{n_j}\le\sup_{m\ge n_k}x_m=u_{n_k}$, giving $\limsup_jx_{n_j}\le\limsup x_n$. The reverse fails: $x_n=(-1)^n$, $n_k=2k-1$: $\limsup x_{2k-1}=-1\lt1$.
IV
Cesàro means and Stolz–Cesàro.
  • 1 Compute $\sigma_n$ for $x_n=(-1)^n$ exactly and show $\sigma_n\to0$.
  • 2 Find $\lim\dfrac{1+\frac12+\cdots+\frac1n}{n}$ and $\lim\dfrac1n\displaystyle\sum_{k=1}^nk^{1/k}$.
  • 3 Find $\lim\dfrac{\sqrt1+\sqrt2+\cdots+\sqrt n}{n^{3/2}}$.
  • 4 Use the geometric-mean form of Theorem 2.19 to show $\left(\displaystyle\prod_{k=1}^n\left(1+\frac1k\right)^k\right)^{1/n}\to e$, and deduce that $\left(\dfrac{(n+1)^n}{n!}\right)^{1/n}\to e$.
  • 5 Let $x_1=1$ and $x_{n+1}=x_n+\dfrac1{x_n}$. Prove $\dfrac{x_n^2}{2n}\to1$, i.e. $x_n\sim\sqrt{2n}$.
  • 6 Show that the block sequence of Interactive 2 ($x_n=1$ if $4^j\le n\lt2\cdot4^j$ for some $j\ge0$, else $0$) has $\liminf\sigma_n=\tfrac13$ and $\limsup\sigma_n=\tfrac23$.
  • 7 Prove: if $(x_n)$ is monotone and $\sigma_n\to L\in\mathbb{R}$, then $x_n\to L$ (a Tauberian theorem).
  • 8 Applied. A sensor reports a reading $x_n$ every second; the display shows the running average $\sigma_n$. The true signal is $5$ V, but the readings carry an error that alternates in sign and decays: $x_n=5+\dfrac{(-1)^n}{\sqrt n}$ volts. Find $|\sigma_n-5|$ to leading order and estimate after how many seconds the display is within $1$ mV of $5$ V; compare with the raw readings.
1 $x_1+\cdots+x_n=-1$ for odd $n$ and $0$ for even $n$, so $\sigma_n=-\tfrac1n$ or $0$; $|\sigma_n|\le\tfrac1n\to0$.
2 Both are Cesàro means of convergent sequences: $\tfrac1k\to0$ gives limit $0$; $k^{1/k}\to1$ (§2.1) gives limit $1$.
3 Stolz with $b_n=n^{3/2}$: $\dfrac{\sqrt{n+1}}{(n+1)^{3/2}-n^{3/2}}$. Multiply top and bottom by $(n+1)^{3/2}+n^{3/2}$: the denominator becomes $(n+1)^3-n^3=3n^2+3n+1$, so the quotient is $\dfrac{\sqrt{n+1}\left((n+1)^{3/2}+n^{3/2}\right)}{3n^2+3n+1}\to\dfrac23$. (Compare $\int_0^1\sqrt x\,dx=\tfrac23$.)
4 $(1+\tfrac1k)^k\to e$, so their geometric means tend to $e$ (Theorem 2.19(b)). The product telescopes: $\prod_{k=1}^n\dfrac{(k+1)^k}{k^k}=\dfrac{(n+1)^n}{n!}$, because each $j$ with $2\le j\le n$ appears as $j^{j-1}$ in the numerator and as $j^j$ in the denominator, leaving $\tfrac1j$. So $\left(\tfrac{(n+1)^n}{n!}\right)^{1/n}\to e$, matching Exercise III.4.
5 $x_n\uparrow\infty$ (§2.1, Exercise III.4). $x_{n+1}^2-x_n^2=2+\dfrac1{x_n^2}\to2$, so by Stolz–Cesàro (or Cesàro applied to the differences) $\dfrac{x_n^2}{n}\to2$. Numerically $x_n^2/(2n)=1.0000033$ at $n=10^6$.
6 Up to $n=2\cdot4^J-1$ the number of ones is $\sum_{j=0}^J4^j=\tfrac{4^{J+1}-1}{3}$, so $\sigma_n=\tfrac{(4^{J+1}-1)/3}{2\cdot4^J-1}\to\tfrac23$; at $n=4^{J+1}-1$ (end of the following block of zeros) $\sigma_n=\tfrac{(4^{J+1}-1)/3}{4^{J+1}-1}=\tfrac13$. Between these indices $\sigma_n$ moves monotonically, so these are the extreme values: $\liminf\sigma_n=\tfrac13$, $\limsup\sigma_n=\tfrac23$.
7 Say $(x_n)$ increases. Then $\sigma_n\le x_n$. If $(x_n)$ were unbounded, $x_n\to\infty$ and $\sigma_n\to\infty$ (Theorem 2.19); so $x_n\to M$ for some real $M$, and then $\sigma_n\to M$, so $M=L$.
8 $\sigma_n-5=\tfrac1n\sum_{k\le n}\tfrac{(-1)^k}{\sqrt k}$. The alternating sum converges (to $\approx-0.6049$; §2.3), so $|\sigma_n-5|\approx\tfrac{0.605}{n}$ V, which is below $10^{-3}$ V once $n\gtrsim605$ — about $10$ minutes. The raw readings need $\tfrac1{\sqrt n}\lt10^{-3}$, i.e. $n\gt10^6$ seconds (over $11$ days). Averaging converts the slow $n^{-1/2}$ error into an $n^{-1}$ error.
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

E1. CSIR Part B Let $x_n=(-1)^n\left(2+\dfrac3n\right)$. Then (A) $\limsup x_n=\sup x_n=\tfrac72$; (B) $\limsup x_n=2$ and $\sup x_n=\tfrac72$; (C) $\limsup x_n=\sup x_n=2$; (D) $\limsup x_n=2$ and $\liminf x_n=-5$.

Idea $\limsup$ looks at tails, $\sup$ at all terms.
Work Even $n$: $2+\tfrac3n\downarrow2$, largest at $n=2$: $\tfrac72$. Odd $n$: $-2-\tfrac3n\uparrow-2$, smallest at $n=1$: $-5$. So $\limsup=2$, $\sup=\tfrac72$, $\liminf=-2$, $\inf=-5$. (A), (C) confuse $\sup$ with $\limsup$; (D) confuses $\liminf$ with $\inf$.
Answer (B).

E2. CSIR Part B GATE MA $\displaystyle\lim_{n\to\infty}\left(\frac{(2n)!}{n!\,n^n}\right)^{1/n}$ equals (A) $\dfrac4e$; (B) $4$; (C) $\dfrac2e$; (D) $\dfrac e4$.

Idea Theorem 2.18: compute the limit of the ratio instead.
Work With $a_n=\tfrac{(2n)!}{n!\,n^n}$: $$\frac{a_{n+1}}{a_n}=\frac{(2n+2)(2n+1)}{n+1}\cdot\frac{n^n}{(n+1)^{n+1}}=\frac{2(2n+1)}{n+1}\Big(\frac{n}{n+1}\Big)^n\to\frac4e .$$ (Numerically, $n=3000$ gives $1.4717$; $4/e=1.4715$.)
Answer (A).

E3. CSIR Part B $\displaystyle\lim_{n\to\infty}\frac1n\sum_{k=1}^n\left(\frac{k}{k+1}\right)^k$ equals (A) $0$; (B) $\dfrac1e$; (C) $1$; (D) $e$.

Idea A Cesàro mean of a convergent sequence.
Work $\left(\tfrac{k}{k+1}\right)^k=\dfrac1{(1+1/k)^k}\to\dfrac1e$, so by Theorem 2.19 the means tend to $\tfrac1e$. The tempting answer $1$ comes from "$\tfrac k{k+1}\to1$", forgetting the exponent.
Answer (B).

E4. CSIR Part C Let $(x_n)$ be a bounded real sequence. Which of the following are true? (One or more options may be correct.) (A) If every convergent subsequence of $(x_n)$ converges to $0$, then $x_n\to0$. (B) If $x_{2n}\to L$ and $x_{3n}\to L$, then $x_n\to L$. (C) $\limsup x_n\le\sup_nx_n$. (D) If $\limsup x_n=\liminf x_n$, then $(x_n)$ is eventually constant.

(A) True Theorem 2.14(a) — boundedness is assumed.
(B) False $x_n=1$ for primes $n\ge5$, $0$ otherwise (Example 2(c)).
(C) True $\limsup x_n=\inf_nu_n\le u_1=\sup_nx_n$.
(D) False Equality means convergence (Theorem 2.16), e.g. $x_n=\tfrac1n$, which is never constant.
Answer (A) and (C).

E5. CSIR Part C Let $(a_n)$, $(b_n)$ be bounded real sequences. Which of the following are true? (A) $\limsup(a_n+b_n)=\limsup a_n+\limsup b_n$. (B) $\liminf a_n+\limsup b_n\le\limsup(a_n+b_n)$. (C) $\limsup(a_nb_n)\le\limsup a_n\cdot\limsup b_n$. (D) $\limsup(-a_n)=-\liminf a_n$.

(A) False $a_n=(-1)^n$, $b_n=-a_n$: $0\ne2$.
(B) True Theorem 2.17(4) (third inequality).
(C) False Without signs: $a_n=b_n=-2,0,-2,0,\dots$ give $4\not\le0$ (Example 4(c)).
(D) True Theorem 2.17(1).
Answer (B) and (D).

E6. CSIR Part C Let $x_n\gt0$ for all $n$. Which of the following are true? (A) If $\dfrac{x_{n+1}}{x_n}\to L$, then $x_n^{1/n}\to L$. (B) If $x_n^{1/n}\to L$, then $\dfrac{x_{n+1}}{x_n}\to L$. (C) If $x_n\to L\gt0$, then $(x_1x_2\cdots x_n)^{1/n}\to L$. (D) If $\dfrac{x_1+\cdots+x_n}{n}\to L\in\mathbb{R}$, then $\dfrac{x_n}{n}\to0$.

(A) True Theorem 2.18.
(B) False $x_n=2+(-1)^n$: roots $\to1$, ratios alternate between $3$ and $\tfrac13$.
(C) True Theorem 2.19(b).
(D) True $\dfrac{x_n}n=\sigma_n-\dfrac{n-1}{n}\sigma_{n-1}\to L-L=0$. A trap: it looks too strong, but it only uses the convergence of $\sigma_n$.
Answer (A), (C) and (D).
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Subsequences and Bolzano–Weierstrass. Definition 2.6 ($n_k\ge k$), Theorem 2.11 with Example 1(a), (b) as divergence tests. Theorem 2.12 (draw the "peaks" picture) and both proofs of Bolzano–Weierstrass — the bisection proof is the one students will need again for $\mathbb{R}^n$ and compactness. Theorem 2.14 and Example 2 (let the class find the counterexample for (c) before showing it). Definition 2.7.
  2. lim sup, lim inf and averages. Definition 2.8 with Interactive 1 projected (first $(-1)^n(1+1/n)$: the red line drops from $\tfrac32$ to $1$; then the sin² example with three subsequential limits; then sin n). Theorem 2.15 proved in full, Theorem 2.16, the strategy box and Example 3. Theorem 2.17 (prove the last inequality; leave the rest as reading), Example 4. Theorem 2.18 with Example 5(a). Cesàro and Stolz–Cesàro with Interactive 2 (show the block sequence) and Example 6(a), (c). Finish with E4 or E6.

Discussion questions. (1) Which information about a sequence is lost when we pass to its set of values, and which is kept by its set of subsequential limits? (2) Why does the proof of Bolzano–Weierstrass need the completeness axiom, and where exactly? (3) Can a sequence have exactly one subsequential limit in $\overline{\mathbb{R}}$ and still diverge? (Yes: $n$, with limit $+\infty$; but in $\mathbb{R}$, for bounded sequences, no.)

Misconceptions to watch for. Confusing $\limsup$ with $\sup$ (and "eventually" with "for all"); thinking a subsequence may repeat or reorder terms; believing $\limsup(x+y)=\limsup x+\limsup y$ or the product rule without signs; assuming Cesàro convergence implies convergence; using Theorem 2.18 in the wrong direction; forgetting the hypotheses $b_n\uparrow\infty$ in Stolz–Cesàro.

Suggested homework. I(1, 2, 5, 6), II(1, 2, 5, 6, 9), III(1, 3, 4, 5, 7), IV(2, 3, 5, 7, 8); Exam corner E1, E4, E6.

🔬 Research–Teaching NexusResearch-tutored · optional, about 45 minutes

Research question. How evenly do the points $\{n\alpha\}$ (the fractional parts of $\alpha,2\alpha,3\alpha,\dots$) fill the interval $[0,1)$ — and can such points replace random numbers in numerical integration?

Why it matters. Banks price options, engineers estimate failure probabilities and film studios render light by averaging a function over many sample points ("Monte Carlo"). With random points the error falls like $1/\sqrt N$. Quasi-Monte Carlo methods use carefully spread deterministic points — the simplest being $\{n\alpha\}$ for irrational $\alpha$ — and often achieve errors close to $1/N$. Behind them is a statement about Cesàro means: for $x_n=\mathbf 1[\{n\alpha\}\lt b]$, Weyl's equidistribution theorem says $\sigma_N\to b$.

What to do.

  1. For $\alpha=\sqrt2$, the golden ratio $\varphi$, $\pi$ and $\tfrac17$, compute the proportion of $n\le N$ with $\{n\alpha\}\lt\tfrac13$, for $N=100$ and $N=10\,000$. Which proportions approach $\tfrac13$?
  2. Explain the result for $\alpha=\tfrac17$ using the periodicity of $\{n/7\}$, and compute the limit of the proportion exactly.
  3. Estimate $\int_0^1x^2\,dx$ by the Cesàro mean $\tfrac1N\sum_{n\le N}\{n\sqrt2\}^2$ and by the mean of $N$ pseudo-random numbers squared, for $N=100$ and $10\,000$. Compare the errors.
  4. Why does $\pi$ behave worse than $\sqrt2$ for small $N$? (Hint: $\pi\approx\tfrac{22}7$ and $\tfrac{355}{113}$.)

Tools: Python (or a spreadsheet); use exact integer arithmetic or high precision for large $N$.

Hand in: A short table of your proportions and integration errors, the exact limit for $\alpha=\tfrac17$, and two sentences relating your findings to Cesàro means and to the density of $\{n\alpha\}$ (Example 1(c)).

Going further. Read about the discrepancy of a point set and the Koksma–Hlawka inequality, which bounds the quasi-Monte Carlo error; or about Sobol' and Halton sequences used in computational finance.

Instructor note: with $N=10\,000$ the proportions are $0.3334$ ($\sqrt2$), $0.3334$ ($\varphi$), $0.3272$ ($\pi$) and $0.4286$ ($\tfrac17$). For $\tfrac17$ the values cycle through $0,\tfrac17,\dots,\tfrac67$, and three of the seven ($0,\tfrac17,\tfrac27$) are $\lt\tfrac13$, so the limit is $\tfrac37\approx0.4286$. For $\int_0^1x^2dx=\tfrac13$ the $\sqrt2$-points give errors $\approx3.0\times10^{-3}$ ($N=100$) and $1.2\times10^{-5}$ ($N=10^4$), random points errors of order $10^{-2}$ and $10^{-3}$ (seed-dependent). Good rational approximations of $\alpha$ make $\{n\alpha\}$ cluster for a while.

🧪 Self-check quiz

0/5 correct

Q1. The Bolzano–Weierstrass theorem says that

Q2. $\limsup_{n\to\infty}(-1)^n\left(1+\tfrac1n\right)$ equals

Q3. A real sequence converges if and only if

Q4. The Cesàro means of $x_n=(-1)^n$

Q5. $\displaystyle\lim_{n\to\infty}\frac{n}{(n!)^{1/n}}$ equals