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

The functional equation log⁡xy=log⁡x+log⁡y\log xy = \log x + \log y

§206 30 min

Why the area under 1/t1/t turns products into sums, and why it is essentially the only differentiable function that does.

Why this matters

The property that made logarithms famous, that they turn multiplication into addition, is not part of our definition. In §206 Hardy derives it from the integral in two lines. The two lines hide a pretty geometric fact about the curve y=1/ty = 1/t, and this lesson brings it out. We also see the converse: any differentiable function with this property is a constant multiple of log⁡x\log x, so the integral definition is not an arbitrary choice.

The idea

Take the region under y=1/ty = 1/t between t=1t = 1 and t=xt = x, and stretch it horizontally by a factor yy: every point (t,h)(t, h) moves to (yt,h)(yt, h). The region becomes yy times wider, so its area is multiplied by yy. But it no longer fits under the curve: at ytyt the curve has height 1/(yt)1/(yt), which is 1/y1/y times the old height 1/t1/t. So also squash the region vertically by the factor 1/y1/y. Now it fits exactly under the curve again, between t=yt = y and t=xyt = xy, and its area has been multiplied by y⋅(1/y)=1y \cdot (1/y) = 1.

The hyperbola y=1/ty = 1/t is unchanged by “stretch sideways by yy, squash vertically by yy”, and so are areas. That is the whole content of the functional equation.

Explore Stretch the area and keep it

Try this. Keep x=2x = 2 and move yy from 11 to 33. Each blue strip over [1,x][1, x] has a matching ochre strip over [y,xy][y, xy]. Compare them using the readout. Then change the number of strips.

t1/t
  • strips over [1, x]
  • the same strips × y, over [y, xy]
  • area over [1, y] = log y
Each strip keeps its area when stretched: e.g. the first strip has area 0.22310.2231 in both places (width × y, height ÷ y). In total, area over [y,xy]=[2.5,5][y, xy] = [2.5, 5] is 0.69310.6931 =log⁡x=0.6931= \log x = 0.6931, so log⁡xy=0.9163+0.6931=1.609\log xy = 0.9163 + 0.6931 = 1.609.
What to notice

The ochre strips get wider and lower as yy grows, but each keeps exactly the area of its blue partner. So the area over [y,xy][y, xy] is always log⁡x\log x, and adding the green area over [1,y][1, y], which is log⁡y\log y, gives the area over [1,xy][1, xy], which is log⁡xy\log xy. (If you make xx larger than yy, the two families overlap; the statement is unaffected.)

Precise statements

Theorem The functional equation of the logarithm Hardy: §206 (1)

For all positive xx and yy,

log⁡xy=log⁡x+log⁡y.\log xy = \log x + \log y .
Corollary Quotients and rational powers

For x,y>0x, y > 0: log⁡(x/y)=log⁡x−log⁡y\log(x/y) = \log x - \log y. For every integer nn, log⁡xn=nlog⁡x\log x^n = n \log x. For integers pp and q>0q > 0, if xp/qx^{p/q} denotes the positive qq-th root of xpx^p, then log⁡xp/q=pqlog⁡x\log x^{p/q} = \dfrac{p}{q}\log x.

Theorem Uniqueness Hardy: Examples LXXXIV, 1

If ff is differentiable for x>0x > 0 and f(xy)=f(x)+f(y)f(xy) = f(x) + f(y) for all positive xx and yy, then f(x)=Clog⁡xf(x) = C \log x for some constant CC.

Proofs

Proof log⁡xy=log⁡x+log⁡y\log xy = \log x + \log y
Strategy

Split the area over [1,xy][1, xy] at yy. The first piece is log⁡y\log y by definition. In the second, the substitution t=yut = yu (the stretch in the picture) turns ∫yxy\int_y^{xy} into ∫1x\int_1^x.

Split the integral at yy

log⁡xy=∫1xydtt=∫1ydtt+∫yxydtt=log⁡y+∫yxydtt.\log xy = \int_1^{xy} \frac{dt}{t} = \int_1^{y} \frac{dt}{t} + \int_y^{xy} \frac{dt}{t} = \log y + \int_y^{xy} \frac{dt}{t}.

This holds whatever the order of 11, yy and xyxy, by the usual convention for integrals with reversed limits.

Substitute t=yut = yu in the second integral

Then dt=y dudt = y\,du and dtt=y duyu=duu\dfrac{dt}{t} = \dfrac{y\,du}{yu} = \dfrac{du}{u}: the factor yy from the width cancels the factor 1/y1/y from the height. When tt runs from yy to xyxy, uu runs from 11 to xx.

Conclude

∫yxydtt=∫1xduu=log⁡x,solog⁡xy=log⁡y+log⁡x.\int_y^{xy} \frac{dt}{t} = \int_1^x \frac{du}{u} = \log x, \qquad\text{so}\qquad \log xy = \log y + \log x .
Proof The corollary
Strategy

Everything follows from the functional equation by algebra, using log⁡1=0\log 1 = 0.

Quotients

log⁡(x/y)+log⁡y=log⁡((x/y) y)=log⁡x\log(x/y) + \log y = \log\big((x/y)\, y\big) = \log x.

Positive integer powers

By induction, log⁡xn=log⁡(xn−1⋅x)=(n−1)log⁡x+log⁡x=nlog⁡x\log x^{n} = \log(x^{n-1} \cdot x) = (n-1)\log x + \log x = n \log x.

Zero and negative powers

log⁡x0=log⁡1=0\log x^0 = \log 1 = 0, and log⁡x−n=log⁡(1/xn)=−log⁡xn=−nlog⁡x\log x^{-n} = \log(1/x^n) = -\log x^n = -n\log x.

Rational powers

Let z=xp/q>0z = x^{p/q} > 0, so zq=xpz^q = x^p. Taking logs, qlog⁡z=plog⁡xq \log z = p \log x, so log⁡z=pqlog⁡x\log z = \dfrac{p}{q}\log x.

The uniqueness theorem says that the functional equation, plus differentiability, already pins the logarithm down. Hardy’s argument differentiates the equation in each variable separately.

Proof Uniqueness of the solution
Strategy

Differentiate f(xy)=f(x)+f(y)f(xy) = f(x) + f(y) with respect to xx and to yy and eliminate f′(xy)f'(xy). This forces xf′(x)x f'(x) to be constant, so f′=C/xf' = C/x, and ff is Clog⁡xC\log x plus a constant that the equation forces to be 00.

Differentiate in each variable

Holding yy fixed and differentiating in xx: yf′(xy)=f′(x)y f'(xy) = f'(x). Holding xx fixed: xf′(xy)=f′(y)x f'(xy) = f'(y).

Eliminate f′(xy)f'(xy)

Multiply the first by xx, the second by yy: xf′(x)=xy f′(xy)=yf′(y)x f'(x) = xy\, f'(xy) = y f'(y) for all x,y>0x, y > 0. A quantity that is the same for every xx is a constant: xf′(x)=Cx f'(x) = C.

Integrate

f′(x)=C/x=ddx(Clog⁡x)f'(x) = C/x = \dfrac{d}{dx}(C\log x), so f(x)−Clog⁡xf(x) - C\log x has zero derivative on (0,∞)(0, \infty) and is a constant C′C' (mean value theorem).

The constant is 00

Put x=y=1x = y = 1 in the equation: f(1)=2f(1)f(1) = 2f(1), so f(1)=0f(1) = 0. Since log⁡1=0\log 1 = 0, C′=f(1)−Clog⁡1=0C' = f(1) - C\log 1 = 0. So f=Clog⁡xf = C\log x. (C=0C = 0 gives the trivial solution f≡0f \equiv 0.)

The constant CC corresponds to the choice of base: we shall see in §217 that log⁡10x=log⁡x/log⁡10\log_{10} x = \log x / \log 10, which is Clog⁡xC\log x with C=1/log⁡10C = 1/\log 10.

Explore Doubling adds the same amount each time

Try this. Set x=1x = 1, then 22, then 44 (and, using the slider, 0.50.5 and 0.250.25). Read off the area each time.

log⁡x=∫1xdtt\log x = \int_1^x \frac{dt}{t}
t1/t
  • y = 1/t
  • area = log x
  • inner rectangle
  • outer rectangle
x−1x=0.75  ≤  log⁡x=1.386  ≤  x−1=3\frac{x-1}{x} = 0.75 \;\le\; \log x = 1.386 \;\le\; x - 1 = 3. The curved area lies between the two rectangles over [1, x].
What to notice

Each doubling adds exactly log⁡2≈0.693\log 2 \approx 0.693: log⁡2n=nlog⁡2\log 2^n = n\log 2. The strips over [1,2][1, 2], [2,4][2, 4], [4,8][4, 8] are stretched copies of one another, which also explains the crude bound log⁡2n≥n/2\log 2^n \ge n/2 of the last lesson.

Worked examples and exercises

Exercise ●●○ Hardy: Examples LXXXIII, 2 (v)

Prove that 22x+1<log⁡x+1x<2x+12x(x+1)\dfrac{2}{2x + 1} < \log\dfrac{x + 1}{x} < \dfrac{2x + 1}{2x(x+1)} for x>0x > 0.

Hint
By the functional equation, log⁡x+1x=log⁡(x+1)−log⁡x=∫xx+1dt/t\log\frac{x+1}{x} = \log(x + 1) - \log x = \int_x^{x+1} dt/t, an area over an interval of length 11.
Hint
The curve 1/t1/t is convex (its second derivative 2/t32/t^3 is positive): it lies below each chord and above each tangent.

Solution

Upper bound. A convex curve lies below its chord. The chord from (x,1/x)(x, 1/x) to (x+1,1/(x+1))(x+1, 1/(x+1)) bounds a trapezium of area 12(1x+1x+1)=2x+12x(x+1)\tfrac12\left(\frac1x + \frac1{x+1}\right) = \dfrac{2x+1}{2x(x+1)}, which therefore exceeds the area under the curve.

Lower bound. A convex curve lies above its tangent. Take the tangent at the midpoint m=x+12m = x + \tfrac12. The area under a straight line over [x,x+1][x, x+1] is its height at the midpoint times the width 11, that is, 1/m=22x+11/m = \dfrac{2}{2x+1}. The curve lies above the tangent, so its area is larger.

Both inequalities are strict because curve and line meet at only one or two points. These are the midpoint and trapezium rules of Chapter VII, and the inequality says the true value lies between them.

Exercise ●●○ Hardy: Examples LXXXIV, 2

Show that a differentiable function ff satisfying

f(x)+f(y)=f ⁣(x+y1−xy)(xy≠1)f(x) + f(y) = f\!\left(\frac{x + y}{1 - xy}\right) \qquad (xy \ne 1)

must be Carctan⁡xC \arctan x for some constant CC.

Hint
Write z=(x+y)/(1−xy)z = (x+y)/(1 - xy) and compute ∂z/∂x\partial z/\partial x and ∂z/∂y\partial z/\partial y.
Hint
You should find ∂z/∂x=(1+y2)/(1−xy)2\partial z/\partial x = (1 + y^2)/(1 - xy)^2. Differentiate the equation in xx and in yy and eliminate f′(z)f'(z).

Solution

Let z=(x+y)/(1−xy)z = (x+y)/(1-xy). Then

∂z∂x=(1−xy)+y(x+y)(1−xy)2=1+y2(1−xy)2,∂z∂y=1+x2(1−xy)2.\frac{\partial z}{\partial x} = \frac{(1 - xy) + y(x + y)}{(1-xy)^2} = \frac{1 + y^2}{(1-xy)^2}, \qquad \frac{\partial z}{\partial y} = \frac{1 + x^2}{(1-xy)^2}.

Differentiating the equation in xx and in yy:

f′(x)=f′(z) 1+y2(1−xy)2,f′(y)=f′(z) 1+x2(1−xy)2.f'(x) = f'(z)\,\frac{1 + y^2}{(1-xy)^2}, \qquad f'(y) = f'(z)\,\frac{1 + x^2}{(1-xy)^2}.

Hence (1+x2)f′(x)=f′(z)(1+x2)(1+y2)(1−xy)2=(1+y2)f′(y)(1 + x^2) f'(x) = f'(z)\dfrac{(1+x^2)(1+y^2)}{(1-xy)^2} = (1 + y^2) f'(y). As before this common value is a constant CC, so f′(x)=C/(1+x2)f'(x) = C/(1 + x^2) and f(x)=Carctan⁡x+C′f(x) = C\arctan x + C'. Putting x=y=0x = y = 0 gives 2f(0)=f(0)2f(0) = f(0), so f(0)=0f(0) = 0 and C′=0C' = 0. The equation is the addition formula arctan⁡x+arctan⁡y=arctan⁡x+y1−xy\arctan x + \arctan y = \arctan\frac{x+y}{1-xy}, valid near 00.

Exercise Halving the work ●○○ Hardy: §206

Given only log⁡2=0.693147\log 2 = 0.693147 and log⁡3=1.098612\log 3 = 1.098612, find log⁡6\log 6, log⁡12\log 12, log⁡98\log \tfrac{9}{8} and log⁡3\log \sqrt{3}.

Hint
Write each number as a product of powers of 22 and 33.

Solution

log⁡6=log⁡2+log⁡3=1.791759\log 6 = \log 2 + \log 3 = 1.791759. log⁡12=2log⁡2+log⁡3=2.484907\log 12 = 2\log 2 + \log 3 = 2.484907. log⁡98=2log⁡3−3log⁡2=2.197225−2.079442=0.117783\log\frac98 = 2\log 3 - 3\log 2 = 2.197225 - 2.079442 = 0.117783. log⁡3=12log⁡3=0.549306\log\sqrt3 = \frac12\log 3 = 0.549306. This is exactly how tables of logarithms were built: compute the logarithms of the primes and add.

Exercise A wrong base point ●●○ Hardy: §206

Let L(x)=∫axdt/tL(x) = \int_a^x dt/t for a fixed a>0a > 0. Find the functional equation satisfied by LL, and explain why the lower limit 11 is the natural choice.

Hint
L(x)=log⁡x−log⁡aL(x) = \log x - \log a.

Solution

L(x)=∫1x−∫1a=log⁡x−log⁡aL(x) = \int_1^x - \int_1^a = \log x - \log a. Then L(xy)=log⁡x+log⁡y−log⁡a=L(x)+L(y)+log⁡aL(xy) = \log x + \log y - \log a = L(x) + L(y) + \log a. The extra constant log⁡a\log a vanishes only when a=1a = 1. In the picture: stretching the region over [a,x][a, x] by yy gives the region over [ay,xy][ay, xy], which starts at ayay, not at aa or at yy, so the areas no longer add up neatly.

Check yourself

Check yourself Enter a number

Given log⁡2=0.6931\log 2 = 0.6931 and log⁡3=1.0986\log 3 = 1.0986, what is log⁡6\log 6 (to 4 decimal places)?

Why

log⁡6=log⁡(2⋅3)=log⁡2+log⁡3=1.7917\log 6 = \log(2 \cdot 3) = \log 2 + \log 3 = 1.7917, or 1.79181.7918 with the more accurate values 0.693147+1.098612=1.7917590.693147 + 1.098612 = 1.791759. Either answer is accepted.

Check yourself True or false?

log⁡(x+y)=log⁡x+log⁡y\log(x + y) = \log x + \log y for all positive xx, yy.

Why

False. The logarithm turns products into sums. For x=y=1x = y = 1 the left side is log⁡2>0\log 2 > 0 and the right side is 00. The correct statement is log⁡xy=log⁡x+log⁡y\log xy = \log x + \log y.

Check yourself Choose one

In the stretching argument, the region over [1,x][1, x] is stretched horizontally by the factor yy. What happens to the curve’s height over the stretched region?

Why

At the new position ytyt the curve has height 1/(yt)=(1/y)(1/t)1/(yt) = (1/y)(1/t). Width times yy and height divided by yy leave every area unchanged. For a curve like 1/t21/t^2 the height would be divided by y2y^2 and areas would shrink by 1/y1/y, which is why only 1/t1/t gives a function that turns products into sums.

Check yourself Choose one

If we had defined L(x)=∫2xdt/tL(x) = \int_2^x dt/t, then L(xy)L(xy) would equal

Why

L(x)=log⁡x−log⁡2L(x) = \log x - \log 2, so L(x)+L(y)=log⁡xy−2log⁡2=L(xy)−log⁡2L(x) + L(y) = \log xy - 2\log 2 = L(xy) - \log 2. Hence L(xy)=L(x)+L(y)+log⁡2L(xy) = L(x) + L(y) + \log 2. Only the base point 11 gives the clean equation.

Hardy’s language

Summary

Summary
  • log⁡xy=log⁡x+log⁡y\log xy = \log x + \log y: split ∫1xy\int_1^{xy} at yy and substitute t=yut = yu in the second part.
  • Geometrically, stretching sideways by yy and squashing by 1/y1/y maps the hyperbola y=1/ty = 1/t to itself and preserves area.
  • Consequences: log⁡(x/y)=log⁡x−log⁡y\log(x/y) = \log x - \log y and log⁡xr=rlog⁡x\log x^r = r\log x for every rational rr.
  • The differentiable solutions of f(xy)=f(x)+f(y)f(xy) = f(x) + f(y) are exactly Clog⁡xC\log x; the lower limit 11 is what makes the extra constant vanish.