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

Starting again from the series

§223 30 min

Define exp⁡x\exp x as ∑xn/n!\sum x^n/n! and rebuild the whole theory from that: the functional equation by multiplying series, the derivative from a two-line estimate, and exe^x for irrational xx as a definition rather than a theorem.

Why this matters

Hardy built the theory in the order area → logarithm → exponential → series. In §223 he sketches the reverse order: start from the series ∑xn/n!\sum x^n/n!, which converges everywhere and needs no geometry, and derive everything else. This is how most modern analysis courses (and Hardy’s own Chapter X, for complex numbers) do it. Seeing both routes shows which facts are definitions and which are theorems, and that the choice is a matter of convenience, not of truth.

The idea

Call the sum of the series exp⁡x\exp x:

exp⁡x=1+x+x22!+x33!+…(all real x).\exp x = 1 + x + \frac{x^2}{2!} + \frac{x^3}{3!} + \dots \qquad (\text{all real } x).

It converges absolutely for every xx (the ratio of consecutive terms is x/n→0x/n \to 0). Two properties carry the whole theory: exp⁡x⋅exp⁡y=exp⁡(x+y)\exp x \cdot \exp y = \exp(x + y), and exp⁡′=exp⁡\exp' = \exp. The first comes from multiplying the two series and regrouping the products; the picture below shows the regrouping.

Explore Multiplying two exponential series

Try this. Row ii and column jj hold the product xii!⋅yjj!\frac{x^i}{i!}\cdot\frac{y^j}{j!}. Step through the diagonals i+j=ni + j = n and compare each diagonal sum with (x+y)n/n!(x+y)^n/n!. Then make xx or yy negative.

x⁰x¹x²x³x⁴x⁵x⁶y⁰y¹y²y³y⁴y⁵y⁶1.0000.5000.1250.0210.0030.00032 × 10⁻⁵1.0000.5000.1250.0210.0030.00032 × 10⁻⁵0.5000.2500.0630.0100.0010.00011 × 10⁻⁵0.1670.0830.0210.0030.00044 × 10⁻⁵4 × 10⁻⁶0.0420.0210.0050.0010.00011 × 10⁻⁵9 × 10⁻⁷0.0080.0040.0010.00022 × 10⁻⁵2 × 10⁻⁶2 × 10⁻⁷0.0010.0010.00023 × 10⁻⁵4 × 10⁻⁶4 × 10⁻⁷3 × 10⁻⁸
Diagonal n=3n = 3: the highlighted cells add up to 0.5625, and (x+y)33!=0.5625\frac{(x+y)^{3}}{3!} = 0.5625: the same number, by the binomial theorem. All complete diagonals together (n≤6n \le 6) give 4.477539, already close to exey=ex+y=4.481689e^{x}e^{y} = e^{x+y} = 4.481689.
What to notice

Every diagonal adds up to exactly (x+y)n/n!(x+y)^n/n!: on the diagonal i+j=ni + j = n the products are 1n!(ni)xiyn−i\frac{1}{n!}\binom{n}{i}x^i y^{n-i}, the terms of the binomial expansion. The cells far from the corner are tiny, so the grid is dominated by its top-left triangle, and the triangle's total is already close to ex+ye^{x+y}. With negative entries (red) the regrouping is still allowed, because the grid of absolute values also has a finite sum.

Precise statements

Definition The exponential function by its series Hardy: §223 (1)

For every real xx, exp⁡x=∑n=0∞xnn!\exp x = \sum_{n=0}^\infty \dfrac{x^n}{n!}.

Theorem The functional equation Hardy: §223 (2)

exp⁡x⋅exp⁡y=exp⁡(x+y)\exp x \cdot \exp y = \exp(x + y) for all real xx, yy.

Theorem The derivative Hardy: §223 (3)

exp⁡\exp is differentiable, hence continuous, and ddxexp⁡x=exp⁡x\dfrac{d}{dx}\exp x = \exp x. More precisely, for 0<∣h∣<20 < |h| < 2,

exp⁡h−1h=1+ρ(h),∣ρ(h)∣≤∣12h∣1−∣12h∣.\frac{\exp h - 1}{h} = 1 + \rho(h), \qquad |\rho(h)| \le \frac{|\tfrac12 h|}{1 - |\tfrac12 h|}.
Theorem exp agrees with powers of ee Hardy: §223

Let e=exp⁡1=∑1/n!e = \exp 1 = \sum 1/n!. Then exp⁡x=ex\exp x = e^x for every rational xx, where exe^x has its elementary meaning (positive roots of integral powers).

From here there are two ways on.

  • Route A. exp⁡\exp is positive, strictly increasing, and runs from 00 to ∞\infty (Exercise 1), so it has an inverse log⁡\log. Since dy/dx=ydy/dx = y for y=exp⁡xy = \exp x, the inverse satisfies dx/dy=1/ydx/dy = 1/y, and with exp⁡0=1\exp 0 = 1 this gives log⁡y=∫1ydt/t\log y = \int_1^y dt/t. We are back at Hardy’s definition of §205, and the rest of the chapter follows.
  • Route B. Use the last theorem as the definition of exe^x for irrational xx: ex:=exp⁡xe^x := \exp x. Then define log⁡\log as the inverse function and ax=exlog⁡aa^x = e^{x\log a}.

Proofs

Proof The functional equation
Strategy

Multiply the two absolutely convergent series, group the products by i+j=ni + j = n, and recognise each group as a term of the binomial expansion.

Products of absolutely convergent series may be regrouped

Both series converge absolutely, so by the theorem on multiplication of series (§202, Examples LXXXI) their product equals the series ∑ncn\sum_n c_n with cn=∑i+j=nxii!yjj!c_n = \sum_{i + j = n} \frac{x^i}{i!}\frac{y^j}{j!}.

Each cnc_n is a binomial expansion

cn=∑i=0nxiyn−ii! (n−i)!=1n!∑i=0n(ni)xiyn−i=(x+y)nn!.c_n = \sum_{i=0}^{n}\frac{x^i y^{n-i}}{i!\,(n-i)!} = \frac{1}{n!}\sum_{i=0}^{n}\binom{n}{i}x^i y^{n-i} = \frac{(x+y)^n}{n!}.

Conclude

So exp⁡x⋅exp⁡y=∑n(x+y)n/n!=exp⁡(x+y)\exp x \cdot \exp y = \sum_n (x+y)^n/n! = \exp(x + y).

Proof The derivative
Strategy

The functional equation reduces the derivative at any point to the derivative at 00, and that is a direct estimate on a series whose terms are dominated by a geometric series.

Reduce to h→0h \to 0 at the origin

exp⁡(x+h)−exp⁡xh=exp⁡x⋅exp⁡h−1h\dfrac{\exp(x+h) - \exp x}{h} = \exp x \cdot \dfrac{\exp h - 1}{h}, by the functional equation.

Expand

exp⁡h−1h=1+h2!+h23!+⋯=1+ρ(h)\dfrac{\exp h - 1}{h} = 1 + \dfrac{h}{2!} + \dfrac{h^2}{3!} + \dots = 1 + \rho(h), where ρ(h)=∑n≥1hn(n+1)!\rho(h) = \sum_{n \ge 1} \dfrac{h^n}{(n+1)!}.

Bound ρ(h)\rho(h) by a geometric series

Since (n+1)!=2⋅3⋯(n+1)≥2n(n+1)! = 2 \cdot 3 \cdots (n+1) \ge 2^n, we get ∣hn/(n+1)!∣≤∣12h∣n|h^n/(n+1)!| \le |\tfrac12 h|^n, so for ∣h∣<2|h| < 2

∣ρ(h)∣≤∣12h∣+∣12h∣2+⋯=∣12h∣1−∣12h∣→0(h→0).|\rho(h)| \le |\tfrac12 h| + |\tfrac12 h|^2 + \dots = \frac{|\tfrac12 h|}{1 - |\tfrac12 h|} \to 0 \quad (h \to 0).

Conclude

The difference quotient tends to exp⁡x\exp x. A differentiable function is continuous, so exp⁡\exp is continuous too.

Proof exp⁡x=ex\exp x = e^x for rational xx
Strategy

Use the functional equation repeatedly: first for positive integers, then for fractions by taking roots, then for negative numbers by taking reciprocals.

Positive integers

By induction, exp⁡(x1+⋯+xn)=exp⁡x1⋯exp⁡xn\exp(x_1 + \dots + x_n) = \exp x_1 \cdots \exp x_n. Taking every xk=1x_k = 1: exp⁡n=en\exp n = e^n.

Positive fractions m/nm/n

Taking every xk=m/nx_k = m/n (nn of them): (exp⁡mn)n=exp⁡m=em\bigl(\exp\frac mn\bigr)^n = \exp m = e^m. Since exp⁡mn>0\exp\frac mn > 0 (its series has positive terms), it is the positive nnth root of eme^m, which is what em/ne^{m/n} means.

Negative rationals

exp⁡x⋅exp⁡(−x)=exp⁡0=1\exp x \cdot \exp(-x) = \exp 0 = 1, so exp⁡(−x)=1/exp⁡x=1/ex=e−x\exp(-x) = 1/\exp x = 1/e^x = e^{-x} for rational x>0x > 0.

Worked examples and exercises

Exercise exp is a bijection onto (0,∞)(0, \infty) ●●○ Hardy: §223

Using only the definition and the two theorems, show that exp⁡x>0\exp x > 0 for all xx, that exp⁡\exp is strictly increasing, and that exp⁡x→∞\exp x \to \infty as x→∞x \to \infty and exp⁡x→0\exp x \to 0 as x→−∞x \to -\infty.

Hint
For x≥0x \ge 0 every term of the series is non-negative. For x<0x < 0 use exp⁡x⋅exp⁡(−x)=1\exp x \cdot \exp(-x) = 1.

Solution

For x≥0x \ge 0, exp⁡x≥1+x≥1\exp x \ge 1 + x \ge 1. For x<0x < 0, exp⁡x=1/exp⁡(−x)\exp x = 1/\exp(-x) with exp⁡(−x)≥1\exp(-x) \ge 1, so 0<exp⁡x≤10 < \exp x \le 1. Hence exp⁡x>0\exp x > 0 everywhere, and since exp⁡′=exp⁡>0\exp' = \exp > 0, exp⁡\exp is strictly increasing (mean value theorem). As x→∞x \to \infty, exp⁡x≥1+x→∞\exp x \ge 1 + x \to \infty; and exp⁡(−x)=1/exp⁡x→0\exp(-x) = 1/\exp x \to 0. Being continuous, exp⁡\exp therefore takes every positive value exactly once (§110), and its inverse log⁡\log is defined on (0,∞)(0, \infty).

Exercise eh≥1+he^h \ge 1 + h ●○○ Hardy: §223

Show that exp⁡h≥1+h\exp h \ge 1 + h for every real hh, with equality only at h=0h = 0.

Hint
For h≥0h \ge 0 read it off the series. For h<0h < 0 consider g(h)=exp⁡h−1−hg(h) = \exp h - 1 - h and its derivative.

Solution

For h>0h > 0 the series gives exp⁡h=1+h+h22!+⋯>1+h\exp h = 1 + h + \frac{h^2}{2!} + \dots > 1 + h. For h<0h < 0, g(h)=exp⁡h−1−hg(h) = \exp h - 1 - h has g′(h)=exp⁡h−1<0g'(h) = \exp h - 1 < 0 (since exp⁡\exp is increasing and exp⁡0=1\exp 0 = 1), so gg decreases on (−∞,0](-\infty, 0] to g(0)=0g(0) = 0; hence g(h)>0g(h) > 0 for h<0h < 0. (The series alone cannot do this: for h<0h < 0 its terms alternate in sign.)

Exercise Only one function reproduces itself ●●○ Hardy: §223; cf. §219 remark

Suppose ff is differentiable on R\mathbb R with f′=ff' = f and f(0)=1f(0) = 1. Show that f=exp⁡f = \exp.

Hint
Differentiate f(x)exp⁡(−x)f(x)\exp(-x).

Solution

Let g(x)=f(x)exp⁡(−x)g(x) = f(x)\exp(-x). Then g′(x)=f′(x)exp⁡(−x)−f(x)exp⁡(−x)=0g'(x) = f'(x)\exp(-x) - f(x)\exp(-x) = 0, so gg is constant, equal to g(0)=1g(0) = 1. Hence f(x)=1/exp⁡(−x)=exp⁡xf(x) = 1/\exp(-x) = \exp x. This is the precise form of Hardy’s remark that the exponential series is the only power series that reproduces itself under differentiation: here not even a power series is assumed.

Exercise The binomial series developed the same way ●●● Hardy: §223, Example

For −1<x<1-1 < x < 1 and real mm let f(m,x)=∑n≥0(mn)xnf(m, x) = \sum_{n \ge 0}\binom{m}{n}x^n. Show that f(m,x) f(m′,x)=f(m+m′,x)f(m, x)\,f(m', x) = f(m + m', x), and deduce that f(p/q,x)f(p/q, x) is the positive qqth root of (1+x)p(1+x)^p for integers pp and q>0q > 0.

Hint
Multiplying the series reduces the first claim to Vandermonde’s identity ∑k(mk)(m′n−k)=(m+m′n)\sum_{k}\binom{m}{k}\binom{m'}{n-k} = \binom{m+m'}{n}. Both sides are polynomials in mm and m′m'.
Hint
To see that f(m,x)>0f(m, x) > 0: it is continuous in xx, equals 11 at x=0x = 0, and never vanishes, because f(m,x) f(−m,x)=f(0,x)=1f(m, x)\,f(-m, x) = f(0, x) = 1.

Solution

Functional equation. Both series converge absolutely for ∣x∣<1|x| < 1, so their product is ∑ncnxn\sum_n c_n x^n with cn=∑k=0n(mk)(m′n−k)c_n = \sum_{k=0}^n\binom{m}{k}\binom{m'}{n-k}. For non-negative integers mm, m′m' this equals (m+m′n)\binom{m+m'}{n}: compare the coefficients of xnx^n in (1+x)m(1+x)m′=(1+x)m+m′(1+x)^m(1+x)^{m'} = (1+x)^{m+m'}. For fixed integer m′≥0m' \ge 0, both sides are polynomials in mm that agree at every non-negative integer, so they agree for all real mm; then, for fixed real mm, the same argument in m′m' gives the identity for all real m,m′m, m'.

Rational exponents. For integers p≥0p \ge 0 the series terminates and f(p,x)=(1+x)pf(p, x) = (1+x)^p; for p<0p < 0, f(p,x)=1/f(−p,x)=(1+x)pf(p, x) = 1/f(-p, x) = (1+x)^p. By the functional equation f(p/q,x)q=f(p,x)=(1+x)pf(p/q, x)^q = f(p, x) = (1+x)^p. Finally f(m,x)≠0f(m, x) \ne 0 (hint 2), and as a power series it is continuous in xx with f(m,0)=1f(m, 0) = 1, so it is positive on (−1,1)(-1, 1) by the intermediate value theorem. Hence f(p/q,x)f(p/q, x) is the positive qqth root of (1+x)p(1+x)^p. For irrational mm one may now define (1+x)m(1+x)^m as f(m,x)f(m, x), exactly as exe^x was defined as exp⁡x\exp x.

Check yourself

Check yourself True or false?

In route B, after proving exp⁡x=ex\exp x = e^x for rational xx, one still has to prove that exp⁡2=e2\exp\sqrt2 = e^{\sqrt2}.

Why

False. In route B, e2e^{\sqrt2} has no meaning until we give it one, and we give it one by defining e2=exp⁡2e^{\sqrt2} = \exp\sqrt2. What makes this a sensible definition is that it agrees with the old meaning at every rational point and that exp⁡\exp is continuous, so it is the only continuous extension.

Check yourself Choose one

Why is absolute convergence needed in the proof of exp⁡x exp⁡y=exp⁡(x+y)\exp x\,\exp y = \exp(x+y)?

Why

The product of two series is a sum over a whole grid of products, and the grid must be summed in some order. For conditionally convergent series, different orders can give different answers, or the diagonal sums may even diverge. Absolute convergence (§202) guarantees that summing by diagonals gives the product of the sums. The binomial theorem (x+y)n=∑(ni)xiyn−i(x+y)^n = \sum\binom{n}{i}x^iy^{n-i} is an algebraic identity valid for all real xx, yy.

Check yourself Enter a number

What upper bound does the estimate ∣ρ(h)∣≤∣12h∣/(1−∣12h∣)|\rho(h)| \le |\tfrac12h|/(1 - |\tfrac12h|) give at h=0.2h = 0.2?

Why

12h=0.1\tfrac12h = 0.1, so the bound is 0.1/0.9=0.111…0.1/0.9 = 0.111\ldots. The true value is ρ(0.2)=(e0.2−1)/0.2−1≈0.1070\rho(0.2) = (e^{0.2} - 1)/0.2 - 1 \approx 0.1070, comfortably inside the bound.

Check yourself Choose one

In route A, how is the logarithm recovered from the series definition of exp⁡\exp?

Why

The series in the first option gives only log⁡(1+x)\log(1 + x) for −1<x≤1-1 < x \le 1. The limit in the third option is a theorem (§216) that itself needs a definition of log⁡\log first. Route A inverts exp⁡\exp, uses the inverse-function rule dx/dy=1/(dy/dx)=1/ydx/dy = 1/(dy/dx) = 1/y, and notes log⁡1=0\log 1 = 0.

Hardy’s language

We shall now give an outline of a method of investigation of the properties of exe^x and log⁡x\log x entirely different in logical order from that followed in the preceding pages.

— G. H. Hardy, §223

Summary

Summary
  • Defining exp⁡x=∑xn/n!\exp x = \sum x^n/n! needs only convergence of a series, no geometry or integrals.
  • exp⁡x exp⁡y=exp⁡(x+y)\exp x\,\exp y = \exp(x+y): multiply absolutely convergent series and sum by diagonals; each diagonal is a binomial expansion.
  • (exp⁡h−1)/h=1+ρ(h)(\exp h - 1)/h = 1 + \rho(h) with ∣ρ(h)∣≤∣12h∣/(1−∣12h∣)|\rho(h)| \le |\tfrac12h|/(1 - |\tfrac12h|), so exp⁡′=exp⁡\exp' = \exp.
  • exp⁡x=ex\exp x = e^x for rational xx; for irrational xx this becomes the definition of exe^x. The logarithm is then the inverse function.
  • The binomial series can be developed in the same way from f(m,x)f(m′,x)=f(m+m′,x)f(m,x)f(m',x) = f(m+m',x).