Mathematics · Book 3 · Bachelor Year 1

University Mathematics — Year 1

University Mathematics — Year 1 · Bachelor Year 1

13Limits and Continuity

The intermediate value theorem and the extreme value theorem were used at High School level on visual faith. With sequences (Chapter 11) and the topology of R\R (Chapter 12) in hand, this chapter proves them — and completes the theory with the monotone bijection theorem (which legitimizes arcsin\arcsin, arcosh\operatorname{arcosh} and their kin from Chapter 4) and Heine’s theorem on uniform continuity.

Throughout, II is an interval and f ⁣:IRf \colon I \to \R; “x0Ix_0 \in \overline I” allows limits at endpoints.

13.1 Limits of functions

Definition 13.1 (Limit at a point)

Let x0Ix_0 \in \overline{I} and R\ell \in \R. Then f(x)f(x) \to \ell as xx0x \to x_0 when

ε>0, δ>0, xI,xx0δ    f(x)ε.\forall \varepsilon > 0,\ \exists \delta > 0,\ \forall x \in I, \qquad \abs{x - x_0} \leq \delta \implies \abs{f(x) - \ell} \leq \varepsilon .

Limits at ±\pm\infty and infinite limits are defined by the same pattern (xx0δ\abs{x - x_0} \leq \delta becomes xMx \geq M; f(x)ε\abs{f(x) - \ell} \leq \varepsilon becomes f(x)Mf(x) \geq M'). One-sided limits restrict xx to x>x0x > x_0 (written xx0+x \to x_0^+) or x<x0x < x_0. The limit is unique when it exists (same proof as Proposition 11.4).

Example 13.2 (A limit through the squeeze)

Compute limx0x1x\lim_{x \to 0} x\,\bigl\lfloor \frac1x \bigr\rfloor. The floor bracketing 1x1<1x1x\frac1x - 1 < \lfloor \frac1x \rfloor \leq \frac1x gives, after multiplying by xx (mind the sign!):

1x<x1x1(x>0),1x1x<1x(x<0),1 - x < x\Bigl\lfloor \frac1x \Bigr\rfloor \leq 1 \quad (x > 0), \qquad 1 \leq x\Bigl\lfloor \frac1x \Bigr\rfloor < 1 - x \quad (x < 0),

and both one-sided squeezes close on 11: the limit is 11. Note what happened: 1x\lfloor \frac1x\rfloor alone has wild jumps near 00, but the factor xx tames each jump (xx times a unit jump is small), and only the bracketing survives. The closing insight: limits of products of a small factor with a bounded-oscillation factor are squeeze problems, never operation-theorem problems — the operations theorem needs both factors to converge.

Theorem 13.3 (Sequential characterization)

f(x)f(x) \to \ell as xx0x \to x_0 if and only if: for every sequence (un)(u_n) of points of II with unx0u_n \to x_0, one has f(un)f(u_n) \to \ell.

Proof. (\Rightarrow) Let unx0u_n \to x_0 and ε>0\varepsilon > 0. Take δ\delta from the definition, then NN with unx0δ\abs{u_n - x_0} \leq \delta for nNn \geq N: beyond NN, f(un)ε\abs{f(u_n) - \ell} \leq \varepsilon.

(\Leftarrow) Contrapositive. If f↛f \not\to \ell: some ε0>0\varepsilon_0 > 0 defeats every δ\delta; choosing δ=1n+1\delta = \frac{1}{n+1} produces unIu_n \in I with unx01n+1\abs{u_n - x_0} \leq \frac{1}{n+1} and f(un)>ε0\abs{f(u_n) - \ell} > \varepsilon_0. Then unx0u_n \to x_0 but f(un)↛f(u_n) \not\to \ell.

Corollary 13.4 (Operations, composition, order)

Sums, products, quotients (nonzero limit downstairs) of limits behave as for sequences; if ff \to \ell at x0x_0 and gmg \to m at \ell, and gg is defined around \ell with g()=mg(\ell) = m or ff \neq \ell near x0x_0, then gfmg \circ f \to m at x0x_0; limits preserve wide inequalities, and the squeeze theorem holds.

Proof. Each statement transfers through Theorem 13.3 to its sequence analogue (Theorems 11.5 and 11.7). For composition, the direct chaining is worth writing once: given ε>0\varepsilon > 0, the limit of gg at \ell provides η>0\eta > 0 with

yη    g(y)mε(y in the domain of g),\abs{y - \ell} \leq \eta \implies \abs{g(y) - m} \leq \varepsilon \quad (y \text{ in the domain of } g),

where the case y=y = \ell is covered because g()=mg(\ell) = m; then the limit of ff at x0x_0 provides δ>0\delta > 0 with xx0δ    f(x)η\abs{x - x_0} \leq \delta \implies \abs{f(x) - \ell} \leq \eta; chaining, xx0δ\abs{x - x_0} \leq \delta gives g(f(x))mε\abs{g(f(x)) - m} \leq \varepsilon. In the alternative hypothesis (ff \neq \ell near x0x_0), the value y=y = \ell is never fed to gg and its value there is irrelevant.

Example 13.5 (Why the composition proviso exists)

Let g(y)=0g(y) = 0 for y0y \neq 0 and g(0)=1g(0) = 1, and let ff be identically 00. Then f(x)0f(x) \to 0 as x0x \to 0, and g(y)0g(y) \to 0 as y0y \to 0; yet g(f(x))=g(0)=1g(f(x)) = g(0) = 1 for every xx, so gf10g \circ f \to 1 \neq 0. The inner function sits exactly on the forbidden value =0\ell = 0 forever, and gg’s limit at 00 ignores what gg does at 00. The proviso of Corollary 13.4 — either g()=mg(\ell) = m (i.e. gg continuous at \ell), or ff \neq \ell near x0x_0 — is precisely what rules this out. The closing insight: in practice one composes continuous functions and the proviso is free; it bites only when limits are taken along punctured neighborhoods, which is why the definition of limxx0\lim_{x \to x_0} used in this book includes the point when it is in the domain.

13.2 Continuity

Definition 13.6

ff is continuous at x0Ix_0 \in I when f(x)f(x0)f(x) \to f(x_0) as xx0x \to x_0; continuous on II when it is continuous at every point. By Theorem 13.3: ff is continuous at x0x_0 iff f(un)f(x0)f(u_n) \to f(x_0) for every sequence unx0u_n \to x_0 in II.

Example 13.7 (A taxonomy of discontinuities)

Three ways to fail at a point, in increasing severity. Removable: f(x)=sinxxf(x) = \frac{\sin x}{x} on R\R^* has limit 11 at 00; defining f(0)=1f(0) = 1 repairs it — the discontinuity was a hole, not a feature. Jump: x\lfloor x \rfloor at an integer has distinct one-sided limits (n1n - 1 and nn); no choice of value can reconcile them, but both half-limits exist. Essential: sin1x\sin\frac1x at 00 has no one-sided limit at all (Exercise 13.1) — oscillation without settlement. Monotone functions can only produce the middle kind (their one-sided limits always exist, being suprema and infima), which is why their discontinuity sets are at most countable — one rational per jump. Derivatives, by Darboux’s theorem (Exercise 14.10), can only produce the last kind: a function with a jump discontinuity is never the derivative of anything.

Proposition 13.8

Sums, products, quotients (where defined) and compositions of continuous functions are continuous. Polynomials, rational fractions (off their poles), \abs{\,\cdot\,}, exp\exp, ln\ln, the trigonometric and hyperbolic functions and their inverses (Chapter 4) are continuous on their domains.

Proof. Operations: Corollary 13.4. Constants and the identity are continuous directly from the definition (δ=ε\delta = \varepsilon serves the identity, any δ\delta the constants); products of continuous functions being continuous, each monomial akxka_k x^k follows by induction on kk, and sums finish the polynomials; a rational fraction is a quotient of two polynomials, continuous wherever the denominator does not vanish. \abs{\,\cdot\,}: reverse triangle inequality, f(x)f(x0)f(x)f(x0)\bigl|\abs{f(x)} - \abs{f(x_0)}\bigr| \leq \abs{f(x) - f(x_0)}, so the same δ\delta works. For the classical functions we grant continuity here; differentiability (proved in Chapter 14) is stronger.

Example 13.9 (Max and min of continuous functions)

If f,gf, g are continuous, so are max(f,g)\max(f, g) and min(f,g)\min(f, g): no case analysis needed, thanks to the identities

max(f,g)=f+g+fg2,min(f,g)=f+gfg2,\max(f, g) = \frac{f + g + \abs{f - g}}{2}, \qquad \min(f, g) = \frac{f + g - \abs{f - g}}{2},

and the continuity of sums and of \abs{\,\cdot\,} (Proposition 13.8). In particular f+=max(f,0)f^+ = \max(f, 0) and f=max(f,0)f^- = \max(-f, 0) are continuous with f=f+ff = f^+ - f^-: the sign-splitting used for series (Chapter 17) and, at full scale, in the integration theory of the Year 3 volume, costs nothing in regularity.

Theorem 13.10 (Intermediate value theorem)

Let ff be continuous on [a,b]\intcc{a}{b} with f(a)0f(b)f(a) \leq 0 \leq f(b). Then f(c)=0f(c) = 0 for some c[a,b]c \in \intcc{a}{b}. Consequently, a continuous function on an interval takes every value between any two of its values: f(I)f(I) is an interval.

Proof. Dichotomy. Set a0=aa_0 = a, b0=bb_0 = b. Given [an,bn]\intcc{a_n}{b_n} with f(an)0f(bn)f(a_n) \leq 0 \leq f(b_n), let mm be the midpoint: if f(m)0f(m) \leq 0 keep [m,bn]\intcc{m}{b_n}, otherwise keep [an,m]\intcc{a_n}{m}; the sign conditions persist. The sequences (an),(bn)(a_n), (b_n) are adjacent (bnan=ba2nb_n - a_n = \frac{b-a}{2^n}), with common limit cc (Theorem 11.11). By continuity and Theorem 11.7: f(c)=limf(an)0f(c) = \lim f(a_n) \leq 0 and f(c)=limf(bn)0f(c) = \lim f(b_n) \geq 0, so f(c)=0f(c) = 0.

For the consequence: given values f(u)<v<f(w)f(u) < v < f(w), apply the above to xf(x)vx \mapsto f(x) - v on the segment with endpoints uu and ww; thus f(I)f(I) is convex, i.e. an interval (Proposition 10.19).

A continuous function with f(a) < 0 < f(b) must cross the axis: the dichotomy proof of  traps a crossing point between adjacent sequences.
A continuous function with f(a)<0<f(b)f(a) < 0 < f(b) must cross the axis: the dichotomy proof of Theorem 13.10 traps a crossing point between adjacent sequences.

Example 13.11 (One equation, the full protocol)

Solve ex=3x\eu^x = 3 - x over R\R: existence, uniqueness, location. Set g(x)=ex+x3g(x) = \eu^x + x - 3, continuous. Location and existence: g(0)=2<0g(0) = -2 < 0 and g(1)=e2>0g(1) = \eu - 2 > 0, so the intermediate value theorem plants a solution in (0,1)\intoo{0}{1}. Uniqueness: gg is a sum of the strictly increasing ex\eu^x and x3x - 3, hence strictly increasing on R\R; a strictly monotone function takes each value at most once, so the solution is unique on all of R\R (not just in the interval probed). The protocol — rearrange to g=0g = 0, sign change for existence, monotonicity for uniqueness — settles most “how many solutions” questions in three lines, and Chapter 14’s variation tables extend it to non-monotone gg by cutting R\R into monotone branches.

Example 13.12 (Dichotomy as an algorithm)

The proof of Theorem 13.10 computes. Take f(x)=x3+x1f(x) = x^3 + x - 1: f(0)=1<0<1=f(1)f(0) = -1 < 0 < 1 = f(1), so a root lies in (0,1)\intoo{0}{1}. Halving:

f(0.5)=0.375<0,f(0.75)=0.171875>0,f(0.625)=0.130859375<0,f(0.5) = -0.375 < 0, \qquad f(0.75) = 0.171875 > 0, \qquad f(0.625) = -0.130859375 < 0 ,

so the root is successively trapped in (0.5,1)\intoo{0.5}{1}, then (0.5,0.75)\intoo{0.5}{0.75}, then (0.625,0.75)\intoo{0.625}{0.75} (true value: c0.6823c \approx 0.6823). After nn steps the error is at most ba2n\frac{b - a}{2^n}: ten steps give three decimals, twenty give six. The closing insight: the intermediate value theorem is not only an existence statement — its dichotomy proof is a guaranteed, if slow, root-finding algorithm, against which the fast but local Newton method of Chapter 14 should be measured.

Theorem 13.13 (Extreme value theorem)

A continuous function on a segment [a,b]\intcc{a}{b} is bounded and attains its bounds: there are c,d[a,b]c, d \in \intcc{a}{b} with

f(c)=inf[a,b]f,f(d)=sup[a,b]f.f(c) = \inf_{\intcc{a}{b}} f, \qquad f(d) = \sup_{\intcc{a}{b}} f .

Combined with Theorem 13.10: the continuous image of a segment is a segment [f(c),f(d)]\intcc{f(c)}{f(d)}.

Proof. Bounded above: otherwise pick unu_n with f(un)nf(u_n) \geq n. By compactness of the segment (Theorem 12.19), a subsequence uφ(n)x[a,b]u_{\varphi(n)} \to x \in \intcc{a}{b}; continuity gives f(uφ(n))f(x)f(u_{\varphi(n)}) \to f(x), but f(uφ(n))φ(n)+f(u_{\varphi(n)}) \geq \varphi(n) \to +\infty: contradiction.

Sup attained: let M=supfM = \sup f and choose vnv_n with f(vn)>M1n+1f(v_n) > M - \frac{1}{n+1} (Proposition 10.4). Extract vφ(n)d[a,b]v_{\varphi(n)} \to d \in \intcc{a}{b}: then f(d)=limf(vφ(n))=Mf(d) = \lim f(v_{\varphi(n)}) = M by the squeeze. The infimum is handled by f-f.

Example 13.14 (Positive minimum on a segment)

Let ff be continuous on [0,1]\intcc{0}{1} with f(x)>0f(x) > 0 for every xx. Then inff>0\inf f > 0: by the extreme value theorem the infimum is a value f(c)f(c), and f(c)>0f(c) > 0 by hypothesis. So a continuous positive function on a segment is bounded away from 00 — a two-line argument used a dozen times in the coming chapters (denominators under control, step-function framings, error bounds). On a non-compact interval this fails spectacularly: f(x)=xf(x) = x on (0,1]\intoc{0}{1} is continuous and positive with inff=0\inf f = 0, not attained. The closing insight: “positive” upgrades to “uniformly positive” exactly when the domain is compact; every hypothesis of the extreme value theorem is load-bearing.

Remark 13.15 (Common pitfalls around the three theorems)

(i) Continuous images: only segments are robust. The continuous image of an open interval need not be open (xx2x \mapsto x^2 maps (1,1)\intoo{-1}{1} onto [0,1)\intco{0}{1}), the image of a closed set need not be closed (arctan\arctan maps the closed R\R onto the open (π2,π2)\intoo{-\frac\pi2}{\frac\pi2}); but the image of a segment is a segment (Theorem 13.13). (ii) The intermediate value theorem needs an interval: the function x1xx \mapsto \frac1x, continuous on [1,0)(0,1]\intco{-1}{0} \cup \intoc{0}{1}, takes the values 1-1 and 11 yet never vanishes — its domain is two disjoint pieces, and the value 00 falls into the gap; always name the interval on which the theorem is applied. (iii) Uniform continuity is a property of the pair (function, set): x2x^2 is uniformly continuous on every segment yet not on R\R (Example 13.20) — the words “uniformly continuous” without a domain are meaningless. (iv) Continuity of the inverse is not formal: it holds on intervals via monotonicity (Theorem 13.16), but a continuous bijection between unions of intervals can have a discontinuous inverse — the remark after that theorem is there because students quote it without the interval hypothesis.

13.3 Monotone functions and inverse functions

Theorem 13.16 (Monotone bijection theorem)

Let ff be continuous and strictly monotonic on an interval II. Then

  1. ff is a bijection from II onto the interval J=f(I)J = f(I);
  2. the inverse f1 ⁣:JIf^{-1} \colon J \to I is strictly monotonic (in the same direction) and continuous.

Proof. Say ff is strictly increasing. (1) Injectivity is immediate from strict monotonicity; surjectivity onto f(I)f(I) is trivial, and f(I)f(I) is an interval by Theorem 13.10.

(2) f1f^{-1} is strictly increasing: if y<yy < y' in JJ but f1(y)f1(y)f^{-1}(y) \geq f^{-1}(y'), applying the increasing ff gives yyy \geq y', absurd. Continuity of f1f^{-1} at y0=f(x0)Jy_0 = f(x_0) \in J: let ε>0\varepsilon > 0. Suppose first x0x_0 interior to II, and shrink ε\varepsilon so that x0±εIx_0 \pm \varepsilon \in I: their images satisfy f(x0ε)<y0<f(x0+ε)f(x_0 - \varepsilon) < y_0 < f(x_0 + \varepsilon). Take δ=min(y0f(x0ε),f(x0+ε)y0)>0\delta = \min\bigl(y_0 - f(x_0 - \varepsilon),\, f(x_0 + \varepsilon) - y_0\bigr) > 0: for yy0δ\abs{y - y_0} \leq \delta, monotonicity of f1f^{-1} squeezes f1(y)f^{-1}(y) between x0εx_0 - \varepsilon and x0+εx_0 + \varepsilon. If x0x_0 is, say, the left endpoint of II, only x0+εx_0 + \varepsilon is available: then y0=minJy_0 = \min J (monotonicity), every yJy \in J with yy0f(x0+ε)y0y - y_0 \leq f(x_0 + \varepsilon) - y_0 satisfies x0f1(y)x0+εx_0 \leq f^{-1}(y) \leq x_0 + \varepsilon, and the one-sided estimate is exactly continuity at an endpoint; the right endpoint is symmetric. (Note: continuity of f1f^{-1} was not deduced from that of ff by symmetry — it is monotonicity on an interval that does the work.)

Remark 13.17

This theorem is what Definition 4.9 and Proposition 4.21 silently used: arcsin\arcsin, arccos\arccos, arctan\arctan, arsinh\operatorname{arsinh}, … are continuous. A complement (Exercise 13.10): a continuous injective function on an interval is automatically strictly monotonic, so the monotonicity hypothesis costs nothing.

Example 13.18 (Roots of all orders)

For nNn \in \N^*, the function f(x)=xnf(x) = x^n is continuous and strictly increasing on [0,+)\intco{0}{+\infty}, with f(0)=0f(0) = 0 and f(x)+f(x) \to +\infty: its image is [0,+)\intco{0}{+\infty} (Theorem 13.10 for the interval structure). The monotone bijection theorem then delivers, in one stroke, a strictly increasing continuous inverse

xx1/n ⁣:[0,+)[0,+):x \mapsto x^{1/n} \colon \intco{0}{+\infty} \to \intco{0}{+\infty} :

existence, uniqueness and continuity of nn-th roots, with no computation. Compare with Exercise 10.12, which built y\sqrt y by hand from the supremum: one chapter of theory has compressed that page of work into two lines, and the same two lines legitimized arcsin\arcsin, arctan\arctan and arsinh\operatorname{arsinh} before them. The closing insight: a good theorem is stored labor.

13.4 Uniform continuity

Definition 13.19

f ⁣:IRf \colon I \to \R is uniformly continuous when

ε>0, δ>0, x,yI,xyδ    f(x)f(y)ε.\forall \varepsilon > 0,\ \exists \delta > 0,\ \forall x, y \in I, \qquad \abs{x - y} \leq \delta \implies \abs{f(x) - f(y)} \leq \varepsilon .

The point: δ\delta depends only on ε\varepsilon, not on the location in II. Uniform continuity implies continuity; a Lipschitz function (f(x)f(y)kxy\abs{f(x) - f(y)} \leq k \abs{x - y}) is uniformly continuous (δ=ε/k\delta = \varepsilon/k).

Example 13.20

xx2x \mapsto x^2 is continuous on R\R but not uniformly continuous: (n+1n)2n2=2+1n22\abs{(n + \frac 1n)^2 - n^2} = 2 + \frac{1}{n^2} \geq 2 although the arguments are at distance 1n0\frac 1n \to 0. On any bounded interval it is Lipschitz, hence uniformly continuous — consistent with Heine’s theorem below.

Example 13.21 (Uniform moduli, explicitly)

On a segment, Heine guarantees a uniform δ\delta; often one can also compute it. For f(x)=x2f(x) = x^2 on [0,10]\intcc{0}{10}:

x2y2=x+yxy20xy,\abs{x^2 - y^2} = \abs{x + y}\,\abs{x - y} \leq 20\,\abs{x - y},

so δ=ε20\delta = \frac{\varepsilon}{20} works uniformly (a Lipschitz modulus, linear in ε\varepsilon). For x\sqrt x on [0,1]\intcc{0}{1}: xyxy\abs{\sqrt x - \sqrt y} \leq \sqrt{\abs{x - y}} (Exercise 13.9), so δ=ε2\delta = \varepsilon^2 works — uniform but not linear: near 00 the square root is steep, and the price appears in the exponent of ε\varepsilon, not in a failure of uniformity. The closing insight: uniform continuity is a spectrum, not a yes/no — the function δ(ε)\delta(\varepsilon), called the modulus, measures how expensive the uniformity is, and Lipschitz is simply its best grade.

Theorem 13.22 (Heine)

A continuous function on a segment [a,b]\intcc{a}{b} is uniformly continuous.

Proof. By contradiction: suppose some ε0>0\varepsilon_0 > 0 defeats every δ\delta. With δ=1n+1\delta = \frac{1}{n+1}, pick xn,yn[a,b]x_n, y_n \in \intcc{a}{b} with xnyn1n+1\abs{x_n - y_n} \leq \frac{1}{n+1} and f(xn)f(yn)>ε0\abs{f(x_n) - f(y_n)} > \varepsilon_0. Compactness (Theorem 12.19) extracts xφ(n)c[a,b]x_{\varphi(n)} \to c \in \intcc{a}{b}; then yφ(n)cy_{\varphi(n)} \to c too (squeeze on xy\abs{x - y}). Continuity at cc gives f(xφ(n))f(c)f(x_{\varphi(n)}) \to f(c) and f(yφ(n))f(c)f(y_{\varphi(n)}) \to f(c), so f(xφ(n))f(yφ(n))0<ε0\abs{f(x_{\varphi(n)}) - f(y_{\varphi(n)})} \to 0 < \varepsilon_0: contradiction.

Example 13.23 (Bounded, continuous, yet not uniformly so)

The function f(x)=sin(x2)f(x) = \sin(x^2) is continuous and bounded on R\R, but not uniformly continuous. Take

xn=2πn,yn=2πn+π2:ynxn=π/22πn+π2+2πn0,x_n = \sqrt{2\pi n}, \qquad y_n = \sqrt{2\pi n + \tfrac\pi2}: \qquad y_n - x_n = \frac{\pi/2}{\sqrt{2\pi n + \frac\pi2} + \sqrt{2\pi n}} \longrightarrow 0 ,

yet f(yn)f(xn)=sin(2πn+π2)sin(2πn)=10=1f(y_n) - f(x_n) = \sin\bigl(2\pi n + \frac\pi2\bigr) - \sin(2\pi n) = 1 - 0 = 1 for every nn: no single δ\delta can serve ε=12\varepsilon = \frac12 everywhere. Geometrically, the oscillations of sin(x2)\sin(x^2) accelerate: the graph completes a full wave over shorter and shorter windows, so the horizontal scale a given ε\varepsilon requires shrinks to zero as xx grows. The closing insight: boundedness does not buy uniformity (this example), and unboundedness does not preclude it (x\sqrt x, Exercise 13.9); what decides is the modulus of oscillation, and Heine’s theorem says compact domains discipline it automatically.

Example 13.24 (The calculator experiment, explained)

Type any number into a calculator and press cos\cos repeatedly: the display locks onto 0.73908510.7390851\dots Why? After one press the value lies in [1,1]\intcc{-1}{1}, after two in [cos1,1][0.54,1]\intcc{\cos 1}{1} \subseteq \intcc{0.54}{1}, a stable interval for cos\cos. On it, cosacosbsin(1)ab\abs{\cos a - \cos b} \leq \sin(1)\,\abs{a - b} with sin1=0.841<1\sin 1 = 0.841\dots < 1 (the product-to-sum bound of Problem 11.1, or the mean value inequality of Chapter 14): the iteration contracts, so by the error-control step of Method 11.23,

unc(0.842)n2u2c0,\abs{u_n - c} \leq (0.842)^{\,n-2}\,\abs{u_2 - c} \longrightarrow 0 ,

where cc is the unique fixed point cosc=c\cos c = c (Exercise 13.6). About 4040 presses buy three decimals (0.842401030.842^{40} \approx 10^{-3}) — a geometric rate, milder than dichotomy’s 2n2^{-n} per step, but each press costs one keystroke while each dichotomy step costs a full sign evaluation. The closing insight: the fixed-point picture of Chapter 11 and the existence theorems of this chapter are two halves of one story — IVT finds cc, contraction reaches it.

Remark 13.25 (Where these theorems work next)

The three pillars of this chapter each power a later one. The intermediate value theorem feeds every existence-of-solutions argument and the monotone bijection theorem; the extreme value theorem turns optimization problems into theorems (Rolle and the mean value theorem in Chapter 14 start exactly there); Heine’s theorem is the reason continuous functions on segments can be integrated in Chapter 15 — the uniform δ\delta is what makes Riemann sums converge. In the Year 2 volume the same trio reappears in normed vector spaces, with compactness doing the work that segments do here.

Remark 13.26 (Perspectives inside this volume)

Continuity is about to be outranked but never retired. Chapter 14 strengthens it to differentiability and returns the favor (differentiable implies continuous); Chapter 15 rests on it twice, through Heine for the construction and through the fundamental theorem, whose central object xaxfx \mapsto \int_a^x f upgrades a merely continuous ff to a C1C^1 primitive. In Chapter 25, continuity in two variables holds a trap worth previewing: the function xyx2+y2\frac{xy}{x^2 + y^2} (extended by 00) is continuous in xx for each fixed yy and in yy for each fixed xx, yet not continuous at the origin — along the diagonal x=yx = y it is constantly 12\frac12. Separate continuity is strictly weaker than continuity: the sequential characterization survives the move to R2\R^2, but sequences must be allowed to approach from every direction, not only along the axes.

13.5 Exercises

Exercise 13.1

Using the sequential characterization, prove that xsin1xx \mapsto \sin\frac 1x has no limit at 0+0^+ (exhibit two sequences). Does xxsin1xx \mapsto x \sin\frac 1x have one?

Solution

Solution of Exercise 13.1.

Take un=12πn+π/2u_n = \frac{1}{2\pi n + \pi/2} and vn=12πnv_n = \frac{1}{2\pi n}: both tend to 0+0^+, yet sin1un=1\sin\frac{1}{u_n} = 1 and sin1vn=0\sin\frac{1}{v_n} = 0. Two sequences, two different limits of images: by Theorem 13.3, no limit at 0+0^+.

xsin1xx \sin\frac1x: squeezed by xsin1xx0\abs{x\sin\frac1x} \leq \abs x \to 0, so the limit at 00 exists and equals 00.

Exercise 13.2

Study the continuity on R\R of f(x)=xf(x) = \lfloor x \rfloor and of g(x)=xxg(x) = x - \lfloor x \rfloor, and of h(x)=x+(xx)2h(x) = \lfloor x \rfloor + (x - \lfloor x\rfloor)^2.

Solution

Solution of Exercise 13.2.

f=f = \lfloor\cdot\rfloor is continuous on RZ\R \setminus \Z (locally constant) and discontinuous at each nZn \in \Z: left limit n1n - 1, value nn.

g(x)=xxg(x) = x - \lfloor x\rfloor: same discontinuity points (the identity is continuous, so gg inherits the jumps of ff); at nZn \in \Z, left limit 10=g(n)1 \neq 0 = g(n).

h(x)=x+(xx)2h(x) = \lfloor x\rfloor + (x - \lfloor x\rfloor)^2: on [n,n+1)\intco{n}{n + 1}, h(x)=n+(xn)2h(x) = n + (x - n)^2, continuous there; at x=nx = n, the left limit is (n1)+1=n=h(n)(n-1) + 1 = n = h(n): the jumps cancel. hh is continuous on R\R (and strictly increasing).

Exercise 13.3

Prove that the equation x53x+1=0x^5 - 3x + 1 = 0 has at least three real solutions (evaluate at well-chosen points and apply Theorem 13.10 on three disjoint segments).

Solution

Solution of Exercise 13.3.

P(x)=x53x+1P(x) = x^5 - 3x + 1: P(2)=32+6+1=25<0P(-2) = -32 + 6 + 1 = -25 < 0; P(0)=1>0P(0) = 1 > 0; P(1)=1<0P(1) = -1 < 0; P(2)=27>0P(2) = 27 > 0. Three sign changes on the disjoint segments [2,0]\intcc{-2}{0}, [0,1]\intcc{0}{1}, [1,2]\intcc{1}{2}: by Theorem 13.10, at least three roots. (Being of degree 55, PP has at most five; a variation study would show exactly three.)

Exercise 13.4

Prove that every polynomial of odd degree has a real root.

Solution

Solution of Exercise 13.4.

Let P=a2m+1X2m+1+P = a_{2m+1}X^{2m+1} + \dots with a2m+1>0a_{2m+1} > 0 (else replace PP by P-P). Factoring the dominant term, P(x)=a2m+1x2m+1(1+o(1))P(x) = a_{2m+1} x^{2m+1}\bigl(1 + o(1)\bigr) as x±x \to \pm\infty: so P(x)+P(x) \to +\infty at ++\infty and -\infty at -\infty. Pick aa with P(a)<0P(a) < 0 and bb with P(b)>0P(b) > 0: the intermediate value theorem on [a,b]\intcc{a}{b} provides a root.

Exercise 13.5 ★★

(Fixed point) Let f ⁣:[0,1][0,1]f \colon \intcc{0}{1} \to \intcc{0}{1} be continuous. Prove that ff has a fixed point: f(c)=cf(c) = c for some cc. Illustrate that neither continuity nor the segment can be dropped.

Solution

Solution of Exercise 13.5.

Let g(x)=f(x)xg(x) = f(x) - x, continuous on [0,1]\intcc{0}{1}. Since ff maps into [0,1]\intcc{0}{1}: g(0)=f(0)0g(0) = f(0) \geq 0 and g(1)=f(1)10g(1) = f(1) - 1 \leq 0. By Theorem 13.10, g(c)=0g(c) = 0 for some cc: a fixed point.

Necessity of the hypotheses: on [0,1]\intcc{0}{1}, the discontinuous map f(x)=1f(x) = 1 for x12x \leq \frac12, f(x)=0f(x) = 0 for x>12x > \frac12 has no fixed point; on the interval (0,1)\intoo{0}{1} (not a segment), f(x)=x2f(x) = \frac x2 is continuous into (0,1)\intoo{0}{1} with no fixed point (the candidate 00 is missing); on R\R, f(x)=x+1f(x) = x + 1.

Exercise 13.6 ★★

Prove that the equation cosx=x\cos x = x has exactly one real solution, and that it lies in (0,1)\intoo{0}{1}.

Solution

Solution of Exercise 13.6.

g(x)=cosxxg(x) = \cos x - x is continuous, g(0)=1>0g(0) = 1 > 0, g(1)=cos11<0g(1) = \cos 1 - 1 < 0: a solution exists in (0,1)\intoo{0}{1} (Theorem 13.10). Uniqueness: gg is strictly decreasing on R\R — for x0x \leq 0, g(x)1x>0g(x) \geq 1 - x > 0 has no zero anyway; and g(x)=sinx10g'(x) = -\sin x - 1 \leq 0 with equality only at isolated points (xπ2mod2πx \equiv -\frac\pi2 \bmod 2\pi), so gg is strictly decreasing (Chapter 14; alternatively: on [0,1]\intcc{0}{1}, cos\cos is strictly decreasing and x-x too, so gg is). A strictly monotone function vanishes at most once.

Exercise 13.7 ★★

Let f ⁣:RRf \colon \R \to \R be continuous with f(x)+f(x) \to +\infty as x±x \to \pm\infty. Prove that ff attains a global minimum on R\R. (Reduce to a segment containing a sublevel set.)

Solution

Solution of Exercise 13.7.

Fix M=f(0)+1M = f(0) + 1. There is A>0A > 0 with f(x)Mf(x) \geq M for xA\abs x \geq A (definition of the two infinite limits; take the larger threshold). On the segment [A,A]\intcc{-A}{A}, the extreme value theorem (Theorem 13.13) provides cc with f(c)=inf[A,A]ff(0)f(c) = \inf_{\intcc{-A}{A}} f \leq f(0). For xA\abs x \geq A: f(x)M>f(0)f(c)f(x) \geq M > f(0) \geq f(c). So f(c)f(c) is the global minimum.

Exercise 13.8 ★★

Let f ⁣:RRf \colon \R \to \R be continuous and periodic (of period T>0T > 0). Prove that ff is bounded and attains its bounds, and that there exists cc with f(c+T2)=f(c)f(c + \frac T2) = f(c). (For the second point, study g(x)=f(x+T2)f(x)g(x) = f(x + \frac T2) - f(x) over one period.)

Solution

Solution of Exercise 13.8.

On the segment [0,T]\intcc{0}{T}, ff is bounded and attains its bounds (Theorem 13.13); by periodicity, these are the bounds on all of R\R, still attained.

Let g(x)=f(x+T2)f(x)g(x) = f(x + \frac T2) - f(x), continuous. Then

g(0)+g(T2)=(f(T2)f(0))+(f(T)f(T2))=f(T)f(0)=0:g(0) + g\bigl(\tfrac T2\bigr) = \bigl(f(\tfrac T2) - f(0)\bigr) + \bigl(f(T) - f(\tfrac T2)\bigr) = f(T) - f(0) = 0 :

g(0)g(0) and g(T2)g(\frac T2) have opposite signs (or one vanishes), so the intermediate value theorem on [0,T/2]\intcc{0}{T/2} gives cc with g(c)=0g(c) = 0, i.e. f(c+T2)=f(c)f(c + \frac T2) = f(c).

Exercise 13.9 ★★

Prove that xxx \mapsto \sqrt x is uniformly continuous on [0,+)\intco{0}{+\infty}, although it is not Lipschitz near 00. (Prove and use xyxy\abs{\sqrt x - \sqrt y} \leq \sqrt{\abs{x - y}}.)

Solution

Solution of Exercise 13.9.

First the inequality: for 0yx0 \leq y \leq x,

(y+xy)2=x+2y(xy)x,\bigl(\sqrt y + \sqrt{x - y}\bigr)^2 = x + 2\sqrt{y(x-y)} \geq x,

so xy+xy\sqrt x \leq \sqrt y + \sqrt{x - y}, i.e. xyxy\sqrt x - \sqrt y \leq \sqrt{x - y}. Hence xyxy\abs{\sqrt x - \sqrt y} \leq \sqrt{\abs{x - y}} for all x,y0x, y \geq 0.

Uniform continuity: given ε>0\varepsilon > 0, take δ=ε2\delta = \varepsilon^2; then xyδ\abs{x - y} \leq \delta implies xyδ=ε\abs{\sqrt x - \sqrt y} \leq \sqrt\delta = \varepsilon.

Not Lipschitz near 00: x0x0=1x+\frac{\sqrt x - \sqrt 0}{x - 0} = \frac{1}{\sqrt x} \to +\infty as x0+x \to 0^+, so no constant kk can dominate all difference quotients.

Exercise 13.10 ★★★

Let ff be continuous and injective on an interval II. Prove that ff is strictly monotonic. Hint: if not, there are a<b<ca < b < c with, say, f(b)>f(a)f(b) > f(a) and f(b)>f(c)f(b) > f(c); apply the intermediate value theorem to a value between max(f(a),f(c))\max(f(a), f(c)) and f(b)f(b) on both sides of bb.

Solution

Solution of Exercise 13.10.

Suppose ff is injective, continuous, and not strictly monotonic. Then there are a<b<ca < b < c in II with f(b)f(b) not between f(a)f(a) and f(c)f(c) — indeed, if for all triples the middle value were between the outer ones, ff would be monotonic (compare any two pairs; a short case check). Say f(b)>max(f(a),f(c))f(b) > \max(f(a), f(c)) (the other case is symmetric, replace ff by f-f). Choose vv with max(f(a),f(c))<v<f(b)\max(f(a), f(c)) < v < f(b). By the intermediate value theorem applied on [a,b]\intcc{a}{b} and on [b,c]\intcc{b}{c}, there are u1(a,b)u_1 \in \intoo{a}{b} and u2(b,c)u_2 \in \intoo{b}{c} with f(u1)=v=f(u2)f(u_1) = v = f(u_2): two distinct points with equal images, contradicting injectivity.

Exercise 13.11 ★★★

(Cauchy’s functional equation, continuous case) Let f ⁣:RRf \colon \R \to \R be continuous with f(x+y)=f(x)+f(y)f(x + y) = f(x) + f(y) for all x,yx, y. Prove that f(x)=f(1)xf(x) = f(1)\,x for all xx: first on N\N, Z\Z, Q\Q (by additivity alone), then on R\R by continuity and density (Theorem 10.14).

Solution

Solution of Exercise 13.11.

f(0)=f(0+0)=2f(0)f(0) = f(0+0) = 2f(0) gives f(0)=0f(0) = 0; f(x)=f(x)f(-x) = -f(x) from 0=f(xx)0 = f(x - x). Set α=f(1)\alpha = f(1). Induction: f(n)=nαf(n) = n\alpha for nNn \in \N, then for nZn \in \Z by oddness. For qNq \in \N^*: qf(pq)=f(p)=pαq\,f(\frac pq) = f(p) = p\alpha (add pq\frac pq to itself qq times), so f(pq)=αpqf(\frac pq) = \alpha\frac pq: f=αidf = \alpha\, \mathrm{id} on Q\Q.

Let now xRx \in \R and (rn)(r_n) a sequence of rationals with rnxr_n \to x (density, Theorem 10.14, applied in nested intervals; or rn=nxnr_n = \frac{\lfloor nx\rfloor}{n}). Continuity: f(x)=limf(rn)=limαrn=αxf(x) = \lim f(r_n) = \lim \alpha r_n = \alpha x.

Exercise 13.12 ★★★

Let f ⁣:[0,+)Rf \colon \intco{0}{+\infty} \to \R be continuous with f(x)Rf(x) \to \ell \in \R as x+x \to +\infty. Prove that ff is uniformly continuous on [0,+)\intco{0}{+\infty}. (Cut at a large AA: Heine on [0,A+1]\intcc{0}{A+1}, the limit beyond AA; make the two regimes overlap.)

Solution

Solution of Exercise 13.12.

Let ε>0\varepsilon > 0. By the limit at ++\infty, there is AA with f(x)ε2\abs{f(x) - \ell} \leq \frac\varepsilon2 for xAx \geq A; hence for x,yAx, y \geq A: f(x)f(y)ε\abs{f(x) - f(y)} \leq \varepsilon (no closeness needed).

On the segment [0,A+1]\intcc{0}{A + 1}, Heine’s theorem (Theorem 13.22) gives δ0>0\delta_0 > 0 for this ε\varepsilon; set δ=min(δ0,1)\delta = \min(\delta_0, 1).

Now take any x,y0x, y \geq 0 with xyδ\abs{x - y} \leq \delta, say xyx \leq y. If yA+1y \leq A + 1: both lie in the segment, and the Heine δ0\delta_0 applies. Otherwise y>A+1y > A + 1, and then xy1>Ax \geq y - 1 > A: both lie in [A,+)\intco{A}{+\infty}, where the limit argument applies. In both cases f(x)f(y)ε\abs{f(x) - f(y)} \leq \varepsilon: uniform continuity.

13.6 Problem: Cauchy’s functional equation and its sisters

Problem 13.1

Weekend problem — f(x+y)=f(x)+f(y)f(x + y) = f(x) + f(y): regularity forces linearity, and the portrait of the monsters

Which functions satisfy f(x+y)=f(x)+f(y)f(x + y) = f(x) + f(y) for all real x,yx, y? Cauchy raised the question in 1821; the answer is a paradigm. Exercise 13.11 shows that such an additive function is linear on Q\Q and that full continuity forces f(x)=cxf(x) = cx. This problem sharpens the hypothesis dramatically — continuity at a single point, or monotonicity, or mere boundedness on one small interval, each suffices — then paints the portrait of a hypothetical nonlinear solution (its graph fills the plane), solves the sister equations characterizing ecx\eu^{cx}, clnxc\ln x, xcx^c and cx2cx^2, and closes with Jensen’s equation and the theorem midpoint-convex ++ continuous     \implies convex. Throughout, additive means: f(x+y)=f(x)+f(y)f(x + y) = f(x) + f(y) for all x,yRx, y \in \R.

Part I — Q\Q-linearity, and one point of continuity.

  1. Let ff be additive. From Exercise 13.11, f(r)=f(1)rf(r) = f(1)\,r for rational rr. Prove the finer statement used below: for every xRx \in \R and rQr \in \Q, f(rx)=rf(x)f(rx) = r\,f(x) (ff is Q\Q-linear).
  2. Suppose the additive ff is continuous at one single point x0x_0. Show that ff is continuous everywhere (compute f(x+h)f(x)f(x + h) - f(x) in terms of f(x0+h)f(x0)f(x_0 + h) - f(x_0)), hence f(x)=f(1)xf(x) = f(1)\,x.
  3. Show that an additive function is determined by its restriction to any dense subgroup: if two additive functions agree on Z+2Z\Z + \sqrt2\,\Z and both are continuous, they are equal — while without continuity, prescribing f(1)=0f(1) = 0 and f(2)=1f(\sqrt 2) = 1 is consistent with Q\Q-linearity on the subgroup. Compute f(m+n2)f(m + n\sqrt2) for this prescription.
  4. Let ff be additive and bounded above by MM on some interval [a,b]\intcc{a}{b} with a<ba < b. Show that ff is bounded above on [0,]\intcc{0}{\ell}, =ba\ell = b - a (translate by aa).

Part II — The regularity ladder.

  1. Continuing question 4: using f(t)+f(t)=f()f(\ell - t) + f(t) = f(\ell), show that ff is also bounded below on [0,]\intcc{0}{\ell}: fC\abs f \leq C there.
  2. Show that f(t)Cn\abs{f(t)} \leq \frac{C}{n} for t[0,/n]t \in \intcc{0}{\ell/n}, and deduce that ff is continuous at 00 (oddness handles the left side), hence everywhere (question 2): an additive function bounded on one interval is linear.
  3. Deduce the monotone case: an additive function nondecreasing on some [a,b]\intcc{a}{b} (a<ba < b) is f(x)=cxf(x) = cx with c0c \geq 0.
  4. Assemble the regularity ladder: for an additive ff, the following are equivalent — (a) f(x)=cxf(x) = cx; (b) ff continuous; (c) ff continuous at one point; (d) ff monotone on some nondegenerate interval; (e) ff bounded on some nondegenerate interval. Arrange the implications so that each is either trivial or already proved.
  5. Verify that questions 4–6 consumed only a bound above: an additive function bounded above on one nondegenerate interval is already linear. Deduce the mirror statement for a bound below, and record the strongest form of the ladder’s rung (e) thus obtained.

Part III — Portrait of a monster. Suppose now ff is additive but not linear.

  1. Show there are nonzero reals u,vu, v with f(u)uf(v)v\dfrac{f(u)}{u} \neq \dfrac{f(v)}{v}, and that the vectors (u,f(u))(u, f(u)) and (v,f(v))(v, f(v)) span the plane (their determinant uf(v)vf(u)u f(v) - v f(u) is nonzero).
  2. Show that the graph of ff contains all points

    r(u,f(u))+s(v,f(v)),r,sQ,r\,(u, f(u)) + s\,(v, f(v)), \qquad r, s \in \Q ,

    and deduce that the graph is dense in R2\R^2: for every point (x0,y0)(x_0, y_0) of the plane and every ε>0\varepsilon > 0, some (x,f(x))(x, f(x)) is within ε\varepsilon of it (solve the 2×22\times2 real system, then approximate the real coefficients by rationals).

  3. Deduce from question 11 the full portrait: a nonlinear additive function is unbounded on every nondegenerate interval, discontinuous at every point, monotone on no interval, and its image of any interval is dense in R\R. Reconcile with question 8.
  4. Monsters exist — on a dense subgroup, constructively: on G=Z+2ZG = \Z + \sqrt2\,\Z define f(m+n2)=nf(m + n\sqrt2) = n. Show ff is well defined and additive on GG, and that ff is unbounded on G(0,ε)G \cap \intoo{0}{\varepsilon} for every ε>0\varepsilon > 0 (for fixed NN, only finitely many g=m+n2(0,1)g = m + n\sqrt2 \in \intoo{0}{1} have nN\abs n \leq N; yet G(0,ε)G \cap \intoo{0}{\varepsilon} is infinite). Explain in one paragraph why extending such an ff to all of R\R requires a basis of R\R as a Q\Q-vector space (a Hamel basis), whose existence is an axiom-of-choice matter beyond this volume.

Part IV — The sister equations. All functions here are continuous.

  1. Let f ⁣:RRf \colon \R \to \R be continuous, not identically 00, with f(x+y)=f(x)f(y)f(x + y) = f(x)f(y). Show f(x)=f(x2)20f(x) = f\bigl(\frac x2\bigr)^2 \geq 0, then f>0f > 0 everywhere, then f(x)=ecxf(x) = \eu^{cx} for some cc: the exponentials are exactly the continuous morphisms from (R,+)(\R, +) to (R,×)(\R^*, \times).
  2. Let f ⁣:(0,+)Rf \colon \intoo{0}{+\infty} \to \R be continuous with f(xy)=f(x)+f(y)f(xy) = f(x) + f(y). Show f(x)=clnxf(x) = c\ln x (transport by exp\exp).
  3. Let f ⁣:(0,+)(0,+)f \colon \intoo{0}{+\infty} \to \intoo{0}{+\infty} be continuous with f(xy)=f(x)f(y)f(xy) = f(x)f(y). Show f(x)=xcf(x) = x^c.
  4. Find all continuous f ⁣:RRf \colon \R \to \R with

    f(x+y)=f(x)+f(y)+f(x)f(y).f(x + y) = f(x) + f(y) + f(x)f(y) .

    (Study h=1+fh = 1 + f; treat the degenerate case separately.)

  5. (Parallelogram) Find all continuous f ⁣:RRf \colon \R \to \R with f(x+y)+f(xy)=2f(x)+2f(y)f(x + y) + f(x - y) = 2f(x) + 2f(y): show ff is even, f(0)=0f(0) = 0, f(nx)=n2f(x)f(nx) = n^2 f(x) by induction, then f(x)=f(1)x2f(x) = f(1)\,x^2. (This equation is the fingerprint of quadratic forms — the parallelogram law that detects, in the Year 2 volume, which norms come from an inner product.)

Part V — Jensen and midpoint convexity.

  1. (Jensen’s equation) Let f ⁣:RRf \colon \R \to \R be continuous with f(x+y2)=f(x)+f(y)2f\bigl(\frac{x+y}{2}\bigr) = \frac{f(x) + f(y)}{2}. Show that g=ff(0)g = f - f(0) satisfies g(x2)=g(x)2g\bigl(\frac x2\bigr) = \frac{g(x)}{2}, deduce that gg is additive, and conclude f(x)=cx+df(x) = cx + d.
  2. Suppose now only the inequality: ff continuous with

    f(x+y2)f(x)+f(y)2(x,yR).f\Bigl(\frac{x + y}{2}\Bigr) \leq \frac{f(x) + f(y)}{2} \qquad (x, y \in \R).

    Prove by induction on nn that for all dyadic weights λ=k2n[0,1]\lambda = \frac{k}{2^n} \in \intcc{0}{1}:

    f(λx+(1λ)y)λf(x)+(1λ)f(y).f\bigl(\lambda x + (1 - \lambda)y\bigr) \leq \lambda f(x) + (1 - \lambda) f(y) .
  3. Extend by continuity and density of the dyadics (Exercise 10.8) to every λ[0,1]\lambda \in \intcc{0}{1}: a continuous midpoint-convex function satisfies the full convexity inequality (the notion studied systematically in Chapter 14).
  4. Show continuity cannot be dropped: a nonlinear additive ff satisfies the midpoint equality of question 19 yet no convexity inequality on any interval (question 12). Moral: midpoint convexity is a countable-stage property (dyadics), convexity a continuum one; continuity is the bridge — exactly as in Parts I–II.

Part VI — Last variations and synthesis.

  1. Find all continuous f ⁣:RRf \colon \R \to \R with f(x+y)=f(x)+f(y)+xyf(x + y) = f(x) + f(y) + xy (subtract the particular solution x22\frac{x^2}{2}).
  2. Prove: if f ⁣:RRf \colon \R \to \R is continuous and additive merely on a dense subgroup GG (i.e. f(g+g)=f(g)+f(g)f(g + g') = f(g) + f(g') for g,gGg, g' \in G), then ff is additive on R\R. More generally, two continuous functions that agree on a dense subset of R\R are equal.
  3. Synthesis, one sentence each: (i) state the regularity ladder of question 8 from memory; (ii) explain why “graph dense in the plane” is the right mental image for the failure of regularity; (iii) list the five classical functions characterized in Part IV–V and the single method that caught them all; (iv) name the two places where density of Q\Q (or of the dyadics) in R\R carried the argument, and the place where it could not (question 13).
Solution

Solution of Problem 13.1.

1. For nNn \in \N: f(nx)=nf(x)f(nx) = n f(x) by induction (f((n+1)x)=f(nx)+f(x)f((n+1)x) = f(nx) + f(x)). Also f(0)=2f(0)f(0) = 2f(0) gives f(0)=0f(0) = 0, and 0=f(xx)=f(x)+f(x)0 = f(x - x) = f(x) + f(-x) gives oddness, so f(nx)=nf(x)f(nx) = nf(x) for nZn \in \Z. For r=pqr = \frac pq: qf(pqx)=f(px)=pf(x)q\,f\bigl(\tfrac pq x\bigr) = f(px) = p\,f(x), so f(rx)=rf(x)f(rx) = r f(x): ff is Q\Q-linear.

2. Additivity gives, for all xx and hh:

f(x+h)f(x)=f(h)=f(x0+h)f(x0).f(x + h) - f(x) = f(h) = f(x_0 + h) - f(x_0) .

As h0h \to 0, the right side tends to 00 by continuity at x0x_0; hence f(x+h)f(x)f(x + h) \to f(x): continuity at every xx. Then Exercise 13.11 yields f(x)=f(1)xf(x) = f(1)\,x.

3. Two continuous additive functions are of the form cxcx and cxc'x (question 2); if they agree on the dense subgroup Z+2Z\Z + \sqrt2\,\Z (Exercise 10.9), then cg=cgc\,g = c'g for some g0g \neq 0 in it: c=cc = c', the functions are equal. Without continuity: Q\Q-linearity only ties values at Q\Q-combinations, and 1,21, \sqrt2 are Q\Q-independent (2Q\sqrt2 \notin \Q), so f(1)=0f(1) = 0, f(2)=1f(\sqrt2) = 1 is consistent and forces, on the subgroup,

f(m+n2)=mf(1)+nf(2)=n.f(m + n\sqrt2) = m\,f(1) + n\,f(\sqrt2) = n .

4. For t[0,]t \in \intcc{0}{\ell}: a+t[a,b]a + t \in \intcc{a}{b}, so f(t)=f(a+t)f(a)Mf(a)=:Mf(t) = f(a + t) - f(a) \leq M - f(a) =: M'.

5. For t[0,]t \in \intcc{0}{\ell}, also t[0,]\ell - t \in \intcc{0}{\ell}, and additivity gives f(t)=f()f(t)f()Mf(t) = f(\ell) - f(\ell - t) \geq f(\ell) - M'. Hence fC\abs{f} \leq C on [0,]\intcc{0}{\ell} with C=max(M,f()M)C = \max\bigl(\abs{M'}, \abs{f(\ell) - M'}\bigr).

6. For t[0,/n]t \in \intcc{0}{\ell/n}: nt[0,]nt \in \intcc{0}{\ell} and f(t)=f(nt)nf(t) = \frac{f(nt)}{n} (question 1), so f(t)Cn\abs{f(t)} \leq \frac Cn. Given ε>0\varepsilon > 0, choose n>Cεn > \frac C\varepsilon: for 0hn0 \leq h \leq \frac\ell n, f(h)ε\abs{f(h)} \leq \varepsilon, and for negative hh use f(h)=f(h)f(h) = -f(-h). So f(h)0=f(0)f(h) \to 0 = f(0) as h0h \to 0: continuity at 00, hence everywhere (question 2), hence f(x)=f(1)xf(x) = f(1)x.

7. If ff is nondecreasing on [a,b]\intcc{a}{b}, then f(a)f(x)f(b)f(a) \leq f(x) \leq f(b) there: bounded, so linear by question 6, f(x)=cxf(x) = cx; and c(ba)=f(b)f(a)0c(b - a) = f(b) - f(a) \geq 0 forces c0c \geq 0.

8. (a)\Rightarrow(b)\Rightarrow(c): trivial. (c)\Rightarrow(a): question 2. (a)\Rightarrow(d): a linear function is monotone everywhere. (d)\Rightarrow(e): a monotone function on [a,b]\intcc{a}{b} is bounded there by its endpoint values. (e)\Rightarrow(a): questions 4–6. The five statements are equivalent — the regularity ladder.

9. Question 4 used only the upper bound MM; question 5 derived the lower bound from the upper one via the reflection f(t)=f()f(t)f(t) = f(\ell) - f(\ell - t); question 6 then ran on fC\abs f \leq C. So: additive and bounded above on one nondegenerate interval already implies linear. For a bound below, apply this to f-f (additive, bounded above). Strongest rung (e): a one-sided bound on one interval suffices.

10. If f(x)x\frac{f(x)}{x} were one constant cc for all x0x \neq 0, ff would be linear; so there are nonzero u,vu, v with f(u)uf(v)v\frac{f(u)}u \neq \frac{f(v)}v, i.e. uf(v)vf(u)0u f(v) - v f(u) \neq 0: the determinant of the vectors (u,f(u))(u, f(u)), (v,f(v))(v, f(v)) is nonzero, and they span R2\R^2.

11. For r,sQr, s \in \Q: f(ru+sv)=rf(u)+sf(v)f(ru + sv) = r f(u) + s f(v) (question 1 twice plus additivity), so the graph contains

(ru+sv,  rf(u)+sf(v))=r(u,f(u))+s(v,f(v)).\bigl(ru + sv,\; r f(u) + s f(v)\bigr) = r\,(u, f(u)) + s\,(v, f(v)) .

Given (x0,y0)(x_0, y_0) and ε>0\varepsilon > 0: the 2×22 \times 2 system a(u,f(u))+b(v,f(v))=(x0,y0)a(u, f(u)) + b(v, f(v)) = (x_0, y_0) has a (unique) real solution (a,b)(a, b) since the determinant is nonzero. Pick rationals rnar_n \to a, snbs_n \to b: then rn(u,f(u))+sn(v,f(v))(x0,y0)r_n(u, f(u)) + s_n(v, f(v)) \to (x_0, y_0) coordinatewise, and each of these points lies on the graph: the graph is dense in R2\R^2.

12. Let II be a nondegenerate interval, x0x_0 its midpoint, MM arbitrary: density provides a graph point within min(I2,1)\min\bigl(\frac{\abs I}{2}, 1\bigr) of (x0,M+1)(x_0, M + 1), i.e. xIx \in I with f(x)>Mf(x) > M: unbounded on II, hence (question 8) discontinuous at every point and monotone on no interval; and for any target y0y_0, graph points near (x0,y0)(x_0, y_0) give f(x)f(x) arbitrarily close to y0y_0 with xIx \in I: f(I)f(I) is dense in R\R. This is question 8 read backwards: since all rungs are equivalent, a nonlinear additive function must fail all of them, everywhere.

13. Well defined: m+n2=m+n2m + n\sqrt2 = m' + n'\sqrt2 forces (nn)2=mmZ(n - n')\sqrt2 = m' - m \in \Z, so n=nn = n' (else 2Q\sqrt2 \in \Q) and m=mm = m'. Additivity on GG is then clear coordinate by coordinate. Unboundedness near 0+0^+: fix ε(0,1)\varepsilon \in \intoo{0}{1} and NNN \in \N. For each fixed nn with nN\abs n \leq N, the condition m+n2(0,1)m + n\sqrt2 \in \intoo{0}{1} pins mm inside an interval of length 11: at most one integer mm per nn, so at most 2N+12N + 1 elements of G(0,1)G \cap \intoo{0}{1} have fN\abs{f} \leq N. But G(0,ε)G \cap \intoo{0}{\varepsilon} is infinite (GG is dense, Exercise 10.9); hence it contains some gg with f(g)>N\abs{f(g)} > N: ff is unbounded on every right neighborhood of 00. Extending ff to an additive function on R\R means choosing values coherently on a family of reals that is Q\Q-linearly independent and spans R\R over Q\Q — a Hamel basis; producing one requires the axiom of choice, and no explicit formula can do it: constructively we own the monster only on GG.

14. f(x)=f(x2+x2)=f(x2)20f(x) = f(\frac x2 + \frac x2) = f(\frac x2)^2 \geq 0. If f(x0)=0f(x_0) = 0, then f(x)=f(xx0)f(x0)=0f(x) = f(x - x_0)f(x_0) = 0 for all xx: excluded. So f>0f > 0 and g=lnfg = \ln \circ f is continuous (Proposition 13.8) with g(x+y)=g(x)+g(y)g(x + y) = g(x) + g(y): by Exercise 13.11, g(x)=cxg(x) = cx, so f(x)=ecxf(x) = \eu^{cx}. Conversely each ecx\eu^{cx} works: the continuous morphisms (R,+)(R,×)(\R, +) \to (\R^*, \times) are exactly the exponentials.

15. g(u)=f(eu)g(u) = f(\eu^u) is continuous and g(u+v)=f(euev)=g(u)+g(v)g(u + v) = f(\eu^u \eu^v) = g(u) + g(v): g(u)=cug(u) = cu, and every x>0x > 0 writes x=eux = \eu^u with u=lnxu = \ln x: f(x)=clnxf(x) = c\ln x.

16. h=lnfh = \ln \circ f is continuous on (0,+)\intoo{0}{+\infty} with h(xy)=h(x)+h(y)h(xy) = h(x) + h(y): by question 15, h(x)=clnxh(x) = c\ln x, so f(x)=eclnx=xcf(x) = \eu^{c\ln x} = x^c.

17. x=y=0x = y = 0: f(0)=2f(0)+f(0)2f(0) = 2f(0) + f(0)^2, so f(0)(1+f(0))=0f(0)(1 + f(0)) = 0. If f(0)=1f(0) = -1: setting y=0y = 0, f(x)=f(x)+f(0)+f(x)f(0)=1f(x) = f(x) + f(0) + f(x)f(0) = -1 for all xx: the constant f1f \equiv -1 (which indeed satisfies the equation). Otherwise f(0)=0f(0) = 0; h=1+fh = 1 + f is continuous, h(0)=1h(0) = 1, and

h(x+y)=1+f(x)+f(y)+f(x)f(y)=h(x)h(y):h(x + y) = 1 + f(x) + f(y) + f(x)f(y) = h(x)\,h(y) :

by question 14, h(x)=ecxh(x) = \eu^{cx}, i.e. f(x)=ecx1f(x) = \eu^{cx} - 1 (the case c=0c = 0 giving f0f \equiv 0). Complete list: f1f \equiv -1 and f(x)=ecx1f(x) = \eu^{cx} - 1, cRc \in \R.

18. x=y=0x = y = 0: 2f(0)=4f(0)2f(0) = 4f(0), so f(0)=0f(0) = 0. x=0x = 0: f(y)+f(y)=2f(y)f(y) + f(-y) = 2f(y), so ff is even. y=xy = x: f(2x)=4f(x)f(2x) = 4f(x). Induction using (x,y)(nx,x)(x, y) \to (nx, x):

f((n+1)x)=2f(nx)+2f(x)f((n1)x)=(2n2+2(n1)2)f(x)=(n+1)2f(x).f((n{+}1)x) = 2f(nx) + 2f(x) - f((n{-}1)x) = (2n^2 + 2 - (n-1)^2) f(x) = (n+1)^2 f(x) .

Then f(x)=f(qxq)=q2f(xq)f(x) = f\bigl(q\cdot\frac xq\bigr) = q^2 f\bigl(\frac xq\bigr) gives f(pqx)=p2q2f(x)f\bigl(\frac pq x\bigr) = \frac{p^2}{q^2}f(x): f(r)=r2f(1)f(r) = r^2 f(1) on Q\Q (evenness handles signs). The two continuous functions ff and xf(1)x2x \mapsto f(1)x^2 agree on the dense set Q\Q, hence everywhere (question 24): f(x)=f(1)x2f(x) = f(1)\,x^2; every cx2cx^2 satisfies the equation.

19. g=ff(0)g = f - f(0) is continuous, g(0)=0g(0) = 0, and satisfies Jensen’s equation (the constants cancel). Taking y=0y = 0: g(x2)=g(x)2g\bigl(\frac x2\bigr) = \frac{g(x)}{2}. Then for all x,yx, y:

g(x+y)2=g(x+y2)=g(x)+g(y)2,\frac{g(x + y)}{2} = g\Bigl(\frac{x+y}{2}\Bigr) = \frac{g(x) + g(y)}{2} ,

so gg is additive and continuous: g(x)=cxg(x) = cx (question 2), and f(x)=cx+df(x) = cx + d with d=f(0)d = f(0). All affine functions satisfy Jensen: the list is complete.

20. Induction on nn. For n=0n = 0: λ{0,1}\lambda \in \{0, 1\}, trivial. Assume the inequality for all weights k2n\frac{k}{2^n}. A weight λ=k2n+1\lambda = \frac{k}{2^{n+1}} with kk even reduces to level nn; for k=2j+1k = 2j + 1, λ\lambda is the midpoint of λ1=j2n\lambda_1 = \frac{j}{2^n} and λ2=j+12n\lambda_2 = \frac{j+1}{2^n}. With zi=λix+(1λi)yz_i = \lambda_i x + (1 - \lambda_i)y: λx+(1λ)y=z1+z22\lambda x + (1-\lambda)y = \frac{z_1 + z_2}{2}, so

f(λx+(1λ)y)f(z1)+f(z2)2(λ1+λ2)f(x)+(2λ1λ2)f(y)2=λf(x)+(1λ)f(y).f\bigl(\lambda x + (1{-}\lambda)y\bigr) \leq \frac{f(z_1) + f(z_2)}{2} \leq \frac{(\lambda_1 + \lambda_2)f(x) + (2 - \lambda_1 - \lambda_2)f(y)}{2} = \lambda f(x) + (1 - \lambda)f(y) .

21. Fix x,yx, y. The maps λf(λx+(1λ)y)\lambda \mapsto f(\lambda x + (1 - \lambda)y) and λλf(x)+(1λ)f(y)\lambda \mapsto \lambda f(x) + (1 - \lambda)f(y) are continuous on [0,1]\intcc{0}{1} (composition and algebra, Proposition 13.8). The inequality holds on the dyadic weights, which are dense in [0,1]\intcc{0}{1} (Exercise 10.8); for arbitrary λ\lambda take dyadics λnλ\lambda_n \to \lambda and pass to the limit (Theorem 13.3 and Theorem 11.7): the convexity inequality holds for every λ[0,1]\lambda \in \intcc{0}{1} — midpoint convexity plus continuity equals convexity (the notion of Chapter 14).

22. A nonlinear additive ff satisfies f(x+y2)=f(x)+f(y)2f\bigl(\frac{x+y}{2}\bigr) = \frac{f(x) + f(y)}{2} exactly (question 1 with r=12r = \frac12, then additivity): it is midpoint-convex, even midpoint-affine. If it satisfied the full convexity inequality on some interval [x,y]\intcc{x}{y}, then for λ[0,1]\lambda \in \intcc{0}{1}: f(λx+(1λ)y)max(f(x),f(y))f(\lambda x + (1-\lambda)y) \leq \max(f(x), f(y)) — bounded above on a nondegenerate interval, hence linear by question 9: contradiction. So continuity in question 21 is not a luxury: without it, midpoint convexity controls only the countable dyadic skeleton, and the continuum in between runs wild.

23. p(x)=x22p(x) = \frac{x^2}{2} satisfies p(x+y)=p(x)+p(y)+xyp(x+y) = p(x) + p(y) + xy. If ff is any continuous solution, g=fpg = f - p is continuous and additive, so g(x)=cxg(x) = cx:

f(x)=x22+cx,cR,f(x) = \frac{x^2}{2} + cx , \qquad c \in \R ,

and each of these is a solution: the list is complete.

24. General principle: if u,vu, v are continuous and agree on a dense DRD \subseteq \R, then for xRx \in \R choose dnDd_n \in D with dnxd_n \to x (Proposition 12.11); u(x)=limu(dn)=limv(dn)=v(x)u(x) = \lim u(d_n) = \lim v(d_n) = v(x). Now let ff be continuous and additive on the dense subgroup GG. Fix x,yRx, y \in \R and take gnxg_n \to x, gnyg'_n \to y with gn,gnGg_n, g'_n \in G; then gn+gnx+yg_n + g'_n \to x + y and, by sequential continuity at x+yx + y, xx and yy:

f(x+y)=limf(gn+gn)=lim(f(gn)+f(gn))=f(x)+f(y):f(x + y) = \lim f(g_n + g'_n) = \lim\bigl(f(g_n) + f(g'_n)\bigr) = f(x) + f(y) :

ff is additive on all of R\R (hence linear, by question 2).

25. (i) For additive ff: linear     \iff continuous     \iff continuous at one point     \iff monotone on some interval     \iff bounded (even one-sidedly) on some interval. (ii) Density of the graph in the plane shows the failure is not a local defect but a global explosion: above every subinterval the values smear over all of R\R, so every regularity property fails everywhere at once. (iii) The catch: cxcx, ecx\eu^{cx}, clnxc\ln x, xcx^c, cx2cx^2, and cx+dcx + d — six characterizations, one method: transport the equation to Cauchy’s, prove the Q\Q-skeleton by induction, upgrade to R\R by density plus continuity. (iv) Density of Q\Q (or the dyadics) carried the upgrades in Exercise 13.11 and in question 21; it carried nothing in question 13, because without continuity values do not propagate from a dense set to its closuredensity transfers information only along continuity.