Chapter IX · The logarithmic, exponential and circular functions of a real variable

Scales of infinity

§210 30 min

The ladder of functions log⁡log⁡x≺log⁡x≺xα≺ex≺…\log\log x \prec \log x \prec x^\alpha \prec e^x \prec \dots, how it is built, and why no such ladder can ever be complete.

Why this matters

When you meet a complicated expression such as x2log⁡x/exx^2 \log x / e^{\sqrt x}, the first question is usually: does it tend to 00, to ∞\infty, or to something in between? The answer comes from knowing which factor grows fastest. §210 organises the standard functions into a single ladder, a scale of infinity, on which every rung beats every rung below it. With the scale in your head, most such questions are settled at a glance; the proofs then only confirm what the scale predicts.

The idea

Line up functions that tend to infinity, from slowest to fastest:

…, log⁡log⁡x, …, log⁡x, log⁡x, (log⁡x)2, …, x0.1, x, x, x2, …, ex, ex, …, eex, …\dots,\ \log\log x,\ \dots,\ \sqrt{\log x},\ \log x,\ (\log x)^2,\ \dots,\ x^{0.1},\ \sqrt{x},\ x,\ x^2,\ \dots,\ e^{\sqrt x},\ e^x,\ \dots,\ e^{e^x},\ \dots

The claim is that each one, divided by any function to its right, tends to 00. That is a claim about eventual behaviour only. For a particular xx the actual sizes may be in a very different order. Try it.

Explore When does the ladder start to tell the truth?

Try this. Start at x=10x = 10 and read the ranking. Then increase tt (where x=10tx = 10^t) slowly and watch which functions are still out of place.

rank at this xfunctionvalueHardy’s place
10exe^{x}220262202610 ✓
9x2x^21001008 ✗
8exe^{\sqrt{x}}23.62423.6249 ✗
7xx10107 ✓
6(log⁡x)2(\log x)^25.30195.30194 ✗
5x\sqrt{x}3.16233.16236 ✗
4log⁡x\log x2.30262.30263 ✗
3log⁡x\sqrt{\log x}1.51741.51742 ✗
2x0.1x^{0.1}1.25891.25895 ✗
1log⁡log⁡x\log\log x0.834030.834031 ✓
At x=101x = 10^{1}: 5 pairs are out of Hardy’s order. The last disagreement disappears at about x=1039.1x = 10^{39.1}; beyond that the order never changes again.
What to notice

At x=10x = 10 the ranking is badly scrambled: x0.1x^{0.1} is below log⁡x\sqrt{\log x} and log⁡x\log x, (log⁡x)2(\log x)^2 beats x\sqrt x, and x2x^2 beats exe^{\sqrt x}. Most disagreements disappear by x=1016x = 10^{16}, but (log⁡x)2(\log x)^2 stays ahead of x0.1x^{0.1} until about x=1039x = 10^{39}. From then on the ranking is exactly Hardy's, and it never changes again.

Precise statements

Hardy uses words (“tends to infinity more slowly than”); a symbol makes the scale easier to write. We follow a common modern notation.

Definition Tending to infinity more slowly Hardy: §210

Let f(x)→∞f(x) \to \infty and ϕ(x)→∞\phi(x) \to \infty as x→∞x \to \infty. We say that ff tends to infinity more slowly than ϕ\phi, and write f≺ϕf \prec \phi, if

f(x)ϕ(x)→0(x→∞).\frac{f(x)}{\phi(x)} \to 0 \qquad (x \to \infty).

The relation is transitive: if f≺ϕf \prec \phi and ϕ≺ψ\phi \prec \psi then f/ψ=(f/ϕ)(ϕ/ψ)→0f/\psi = (f/\phi)(\phi/\psi) \to 0.

Theorem log⁡log⁡x\log\log x is slower than every power of log⁡x\log x Hardy: §210

For every α>0\alpha > 0, log⁡log⁡x(log⁡x)α→0\dfrac{\log\log x}{(\log x)^\alpha} \to 0 as x→∞x \to \infty. More generally, each iterated logarithm is slower than every positive power of the one before it: if log⁡1x=log⁡x\log_1 x = \log x and log⁡k+1x=log⁡(log⁡kx)\log_{k+1} x = \log(\log_k x), then log⁡k+1x≺(log⁡kx)α\log_{k+1} x \prec (\log_k x)^\alpha for every α>0\alpha > 0.

Theorem The logarithmic scale Hardy: §210

In the sequence

x, x, x3, …, xn, …, log⁡x, log⁡x, …, log⁡xn, …, log⁡log⁡x, …, log⁡log⁡log⁡x, …x,\ \sqrt{x},\ \sqrt[3]{x},\ \dots,\ \sqrt[n]{x},\ \dots,\ \log x,\ \sqrt{\log x},\ \dots,\ \sqrt[n]{\log x},\ \dots,\ \log\log x,\ \dots,\ \log\log\log x,\ \dots

every function tends to infinity, and every function tends to infinity more slowly than each function to its left.

Hardy writes the scale with the fastest functions on the left. In §213 he extends it at the other end with exe^x, eexe^{e^x}, …, and the two pieces combine into the single ladder displayed under The idea.

Theorem No scale is complete Hardy: Examples LXXXV, 1, 3

If f≺Ff \prec F, then ϕ=fF\phi = \sqrt{fF} satisfies f≺ϕ≺Ff \prec \phi \prec F. In particular, between any two rungs of the scale there is room for another function; and there are functions, such as x/log⁡log⁡x\sqrt{x}/\log\log x, that lie strictly between x\sqrt{x} and all the functions x (log⁡x)−α\sqrt{x}\,(\log x)^{-\alpha}, α>0\alpha > 0, and so fit into no gap formed by the functions written above.

Proofs

Proof log⁡log⁡x≺(log⁡x)α\log\log x \prec (\log x)^\alpha
Strategy

This is the previous lesson’s theorem with log⁡x\log x in place of xx. Substituting y=log⁡xy = \log x turns a statement about log⁡log⁡x\log\log x against powers of log⁡x\log x into one about log⁡y\log y against powers of yy.

Known: log⁡y/yα→0\log y / y^\alpha \to 0 as y→∞y \to \infty

This is §208, valid for every α>0\alpha > 0.

Put y=log⁡xy = \log x

As x→∞x \to \infty, y=log⁡x→∞y = \log x \to \infty. So log⁡log⁡x(log⁡x)α=log⁡yyα→0\dfrac{\log\log x}{(\log x)^\alpha} = \dfrac{\log y}{y^\alpha} \to 0. (Formally: given δ\delta, choose YY so that the ratio is below δ\delta for y>Yy > Y; then it is below δ\delta for x>eYx > e^Y, or, without ee, for every xx with log⁡x>Y\log x > Y.)

Iterate

Applying the same substitution with y=log⁡kxy = \log_k x gives log⁡k+1x/(log⁡kx)α→0\log_{k+1} x / (\log_k x)^\alpha \to 0, for every kk.

Proof The logarithmic scale
Strategy

There are three kinds of neighbours in the sequence: two roots of xx, two roots of log⁡x\log x (or of an iterated log), and the junction where one block ends and the next begins. Each is a ratio that we can show tends to 00.

Every function tends to infinity

x1/n→∞x^{1/n} \to \infty; log⁡kx→∞\log_k x \to \infty for each kk, since log⁡y→∞\log y \to \infty as y→∞y \to \infty; and positive roots of functions tending to ∞\infty tend to ∞\infty.

Within a block

x1/(n+1)x1/n=x−1/(n(n+1))→0\dfrac{x^{1/(n+1)}}{x^{1/n}} = x^{-1/(n(n+1))} \to 0. The same calculation with log⁡kx\log_k x in place of xx handles the other blocks.

At the junction between powers of xx and log⁡x\log x

log⁡x\log x must be slower than every x1/nx^{1/n}: this is log⁡x/x1/n→0\log x / x^{1/n} \to 0 from §208.

At the junctions between log⁡k\log_k and log⁡k+1\log_{k+1}

log⁡k+1x\log_{k+1} x must be slower than every (log⁡kx)1/n(\log_k x)^{1/n}: this is the theorem just proved with α=1/n\alpha = 1/n.

Every pair, not only neighbours

Any function is related to any function to its right by a finite chain of neighbouring steps, and ≺\prec is transitive.

Proof No scale is complete
Strategy

The geometric mean of two functions lies “halfway” between them on the scale, because ratios of geometric means are square roots of ratios.

ϕ=fF\phi = \sqrt{fF} tends to infinity

ϕ≥f⋅f=f\phi \ge \sqrt{f \cdot f} = f eventually (since F≥fF \ge f eventually, as f/F→0f/F \to 0), and f→∞f \to \infty.

f≺ϕf \prec \phi

fϕ=fF→0\dfrac{f}{\phi} = \sqrt{\dfrac{f}{F}} \to 0.

ϕ≺F\phi \prec F

ϕF=fF→0\dfrac{\phi}{F} = \sqrt{\dfrac{f}{F}} \to 0.

A function outside every gap

Let g=x/log⁡log⁡xg = \sqrt x / \log\log x. Then g/x=1/log⁡log⁡x→0g/\sqrt x = 1/\log\log x \to 0, so g≺xg \prec \sqrt x; and for every α>0\alpha > 0, x (log⁡x)−αg=log⁡log⁡x(log⁡x)α→0\dfrac{\sqrt x\,(\log x)^{-\alpha}}{g} = \dfrac{\log\log x}{(\log x)^{\alpha}} \to 0, so gg beats every x (log⁡x)−α\sqrt{x}\,(\log x)^{-\alpha}. No function of the form x (log⁡x)−α\sqrt x\,(\log x)^{-\alpha} lies between gg and x\sqrt x, so gg falls into a gap that the family x (log⁡x)−α\sqrt x\,(\log x)^{-\alpha} cannot fill.

Hardy calls this incompleteness “an inherent characteristic” of the scale: however many functions we add, there remain functions that fit between them. A deeper theorem of du Bois-Reymond, which Hardy treats in his tract Orders of Infinity, shows that no countable scale can be complete.

Worked examples and exercises

Exercise ●○○ Hardy: Examples LXXXV, 2

Find a function that tends to infinity more slowly than x\sqrt x but more quickly than xαx^\alpha for every α<12\alpha < \tfrac12.

Hint
Divide x\sqrt x by something that tends to infinity, but more slowly than any power.

Solution

Take ϕ(x)=x/log⁡x\phi(x) = \sqrt{x}/\log x (or x (log⁡x)−β\sqrt x\,(\log x)^{-\beta} for any β>0\beta > 0). Then ϕ/x=1/log⁡x→0\phi/\sqrt x = 1/\log x \to 0. And for α<12\alpha < \frac12, xα/ϕ=(log⁡x)/x1/2−α→0x^\alpha/\phi = (\log x)/x^{1/2 - \alpha} \to 0 by §208, since 12−α>0\tfrac12 - \alpha > 0. Also ϕ→∞\phi \to \infty for the same reason with α=0\alpha = 0.

Exercise ●●○ Hardy: Examples LXXXV, 4

How does

f(x)=xα(log⁡x)α′(log⁡log⁡x)α′′xβ(log⁡x)β′(log⁡log⁡x)β′′f(x) = \frac{x^{\alpha}(\log x)^{\alpha'}(\log\log x)^{\alpha''}}{x^{\beta}(\log x)^{\beta'}(\log\log x)^{\beta''}}

behave as x→∞x \to \infty?

Hint
Write f=xa(log⁡x)a′(log⁡log⁡x)a′′f = x^{a}(\log x)^{a'}(\log\log x)^{a''} with a=α−βa = \alpha - \beta, and so on. Which factor dominates?

Solution

Let a=α−βa = \alpha - \beta, a′=α′−β′a' = \alpha' - \beta', a′′=α′′−β′′a'' = \alpha'' - \beta'', so f=xa(log⁡x)a′(log⁡log⁡x)a′′f = x^{a}(\log x)^{a'}(\log\log x)^{a''}.

If a>0a > 0: xax^{a} beats any power of log⁡x\log x or log⁡log⁡x\log\log x, so f→∞f \to \infty. Precisely, (log⁡x)a′(log⁡log⁡x)a′′≥(log⁡x)−c(\log x)^{a'}(\log\log x)^{a''} \ge (\log x)^{-c} for some c>0c > 0 and large xx, and xa(log⁡x)−c→∞x^{a}(\log x)^{-c} \to \infty. If a<0a < 0, f→0f \to 0 in the same way.

If a=0a = 0, the power of xx disappears and (log⁡x)a′(\log x)^{a'} dominates: f→∞f \to \infty if a′>0a' > 0, f→0f \to 0 if a′<0a' < 0 (because (log⁡log⁡x)a′′(\log\log x)^{a''} is slower than every power of log⁡x\log x).

If a=a′=0a = a' = 0, then f=(log⁡log⁡x)a′′f = (\log\log x)^{a''}, which tends to ∞\infty, 11 or 00 as a′′>0a'' > 0, =0= 0 or <0< 0.

In short: the first non-zero exponent in the list a,a′,a′′a, a', a'' decides, as in a dictionary ordering.

Exercise ●●○ Hardy: Examples LXXXV, 5

Arrange in order of growth for large xx:

xlog⁡x,xlog⁡xlog⁡log⁡x,xlog⁡log⁡xlog⁡x,xlog⁡log⁡log⁡xlog⁡log⁡x.\frac{x}{\sqrt{\log x}}, \qquad \frac{x\sqrt{\log x}}{\log\log x}, \qquad \frac{x\log\log x}{\sqrt{\log x}}, \qquad \frac{x\log\log\log x}{\sqrt{\log\log x}}.
Hint
All four are xx times a slowly varying factor. Compare the factors using the “first non-zero exponent decides” rule.
Solution

Divide each by xx: the factors are A=(log⁡x)−1/2A = (\log x)^{-1/2}, B=(log⁡x)1/2(log⁡log⁡x)−1B = (\log x)^{1/2}(\log\log x)^{-1}, C=(log⁡x)−1/2log⁡log⁡xC = (\log x)^{-1/2}\log\log x, D=(log⁡log⁡x)−1/2log⁡log⁡log⁡xD = (\log\log x)^{-1/2}\log\log\log x.

B→∞B \to \infty (positive power of log⁡x\log x); the other three tend to 00. DD contains no power of log⁡x\log x, so it tends to 00 more slowly than AA and CC, which both carry (log⁡x)−1/2(\log x)^{-1/2}. Between those, C/A=log⁡log⁡x→∞C/A = \log\log x \to \infty. So the order, slowest first, is

xlog⁡x≺xlog⁡log⁡xlog⁡x≺xlog⁡log⁡log⁡xlog⁡log⁡x≺xlog⁡xlog⁡log⁡x.\frac{x}{\sqrt{\log x}} \prec \frac{x\log\log x}{\sqrt{\log x}} \prec \frac{x\log\log\log x}{\sqrt{\log\log x}} \prec \frac{x\sqrt{\log x}}{\log\log x}.
Exercise ●●○ Hardy: Examples LXXXV, 6

Show that, for large xx, log⁡(x+1)=log⁡x+O(1/x)\log(x + 1) = \log x + O(1/x), 12log⁡x+1x−1=1x+O ⁣(1x2)\tfrac12\log\dfrac{x+1}{x-1} = \dfrac1x + O\!\left(\dfrac1{x^2}\right), and log⁡(xlog⁡x)∼log⁡x\log(x\log x) \sim \log x.

Hint
Use h1+h<log⁡(1+h)<h\dfrac{h}{1+h} < \log(1+h) < h from the first lesson, with h=±1/xh = \pm 1/x.

Solution

First. log⁡(x+1)−log⁡x=log⁡(1+1/x)\log(x+1) - \log x = \log(1 + 1/x), and 0<log⁡(1+1/x)<1/x0 < \log(1 + 1/x) < 1/x. So the difference is O(1/x)O(1/x).

Second. 12log⁡x+1x−1=12{log⁡(1+1/x)−log⁡(1−1/x)}\tfrac12\log\frac{x+1}{x-1} = \tfrac12\{\log(1 + 1/x) - \log(1 - 1/x)\}. With h=1/xh = 1/x: h1+h<log⁡(1+h)<h\frac{h}{1+h} < \log(1+h) < h and h<−log⁡(1−h)<h1−hh < -\log(1-h) < \frac{h}{1-h}. Adding, h1+h+h<log⁡1+h1−h<h+h1−h\frac{h}{1+h} + h < \log\frac{1+h}{1-h} < h + \frac{h}{1-h}, i.e. 2h−h21+h<log⁡1+h1−h<2h+h21−h2h - \frac{h^2}{1+h} < \log\frac{1+h}{1-h} < 2h + \frac{h^2}{1-h}. Halving, the quantity differs from h=1/xh = 1/x by less than h21−h≤2h2=2/x2\frac{h^2}{1-h} \le 2h^2 = 2/x^2 for x≥2x \ge 2. So it is 1/x+O(1/x2)1/x + O(1/x^2).

Third. log⁡(xlog⁡x)=log⁡x+log⁡log⁡x\log(x\log x) = \log x + \log\log x, so log⁡(xlog⁡x)log⁡x=1+log⁡log⁡xlog⁡x→1\dfrac{\log(x\log x)}{\log x} = 1 + \dfrac{\log\log x}{\log x} \to 1.

Exercise ●○○ Hardy: Examples LXXXV, 7

Show that ddxlog⁡log⁡x=1xlog⁡x\dfrac{d}{dx}\log\log x = \dfrac{1}{x\log x}, and hence ∫dxxlog⁡x=log⁡log⁡x\displaystyle\int \frac{dx}{x\log x} = \log\log x (for x>1x > 1). What is ddxlog⁡log⁡log⁡x\dfrac{d}{dx}\log\log\log x?

Hint
Chain rule.

Solution

ddxlog⁡(log⁡x)=1log⁡x⋅1x\frac{d}{dx}\log(\log x) = \frac{1}{\log x}\cdot\frac{1}{x}. Reading this backwards gives the integral. Again by the chain rule, ddxlog⁡log⁡log⁡x=1log⁡log⁡x⋅1xlog⁡x\frac{d}{dx}\log\log\log x = \frac{1}{\log\log x}\cdot\frac{1}{x\log x}, so ∫dxxlog⁡xlog⁡log⁡x=log⁡log⁡log⁡x\int \frac{dx}{x\log x\log\log x} = \log\log\log x (for x>ex > e). These integrals drive the logarithmic tests for series in §218.

Check yourself

Check yourself True or false?

Since log⁡x≺x0.1\log x \prec x^{0.1}, we have log⁡x<x0.1\log x < x^{0.1} for every x>1x > 1.

Why

False. f≺ϕf \prec \phi is a statement about the limit of f/ϕf/\phi, so it says f<ϕf < \phi only for sufficiently large xx. Here log⁡x>x0.1\log x > x^{0.1} for all xx from about e1.12≈3.06e^{1.12} \approx 3.06 up to about 3.4×10153.4\times 10^{15}.

Check yourself Choose one

Which function lies strictly between log⁡x\log x and log⁡x\sqrt{\log x} on the scale (slower than the first, faster than the second)?

Why

(log⁡x)3/4(\log x)^{3/4} is the geometric mean of log⁡x\log x and log⁡x\sqrt{\log x}. log⁡log⁡x\log\log x is slower than both; x0.01x^{0.01} is faster than both. 2log⁡x2\log x is not faster or slower than log⁡x\log x: the ratio tends to 22, not to 00 or ∞\infty. Constant factors are invisible on the scale.

Check yourself Choose one

As x→∞x \to \infty, x2(log⁡x)−5x2(log⁡log⁡x)100\dfrac{x^{2}(\log x)^{-5}}{x^{2}(\log\log x)^{100}}

Why

The powers of xx cancel, so the first non-zero exponent is that of log⁡x\log x, namely −5-5. The quotient is (log⁡x)−5(log⁡log⁡x)−100→0(\log x)^{-5}(\log\log x)^{-100} \to 0. A large power of log⁡log⁡x\log\log x cannot compensate for any power of log⁡x\log x.

Hardy’s language

It will be gathered from these examples that incompleteness is an inherent characteristic of the logarithmic scale of infinity.

— G. H. Hardy, §210

Summary

Summary
  • f≺ϕf \prec \phi (“more slowly”) means f/ϕ→0f/\phi \to 0; it is a statement about large xx only.
  • Substituting x↦log⁡xx \mapsto \log x in log⁡x≺xα\log x \prec x^\alpha gives log⁡log⁡x≺(log⁡x)α\log\log x \prec (\log x)^\alpha, and so on for every iterated logarithm.
  • The scale: ⋯≺log⁡log⁡x≺(log⁡x)α≺xβ≺ex≺eex≺…\dots \prec \log\log x \prec (\log x)^{\alpha} \prec x^{\beta} \prec e^{x} \prec e^{e^x} \prec \dots for any positive α\alpha, β\beta.
  • For products of powers of xx, log⁡x\log x, log⁡log⁡x\log\log x, the first non-zero exponent decides.
  • No scale is complete: fF\sqrt{fF} fits between ff and FF.