Appendices · Appendices

Every equation has a root: the least value of ∣f(z)∣|f(z)|

Appendix II 40 min

Appendix II, proof (B): ∣f(z)∣|f(z)| attains a least value, and at a point where f≠0f \ne 0 one can always go lower, so the least value is 00.

Why this matters

Hardy’s second proof of the fundamental theorem of algebra is the one most textbooks give, and it is a model of a very common kind of argument: a minimum exists, and a minimum cannot be at a bad point, so it is at a good point. It needs two ingredients. One is the two-dimensional version of the theorem of §103 that a continuous function on a closed bounded region attains its bounds. The other is an algebraic observation, often called d’Alembert’s lemma: at any point where f(z0)≠0f(z_0) \ne 0, there are nearby points where ∣f∣|f| is smaller.

(The appendix has no section numbers. Hardy’s proof extends the results of §§103–105 to functions of two variables.)

The idea

Think of ∣f(x+iy)∣|f(x + iy)| as the height of a landscape over the plane. Far out, ∣f(z)∣≈∣a0∣∣z∣n|f(z)| \approx |a_0||z|^n is huge, so the landscape is a giant bowl. A bowl that is continuous and closed off at the rim has a lowest point. The theorem says that the lowest point is at height 00, which is the same as saying that ff has a root there.

Why can the lowest point not be at a positive height? Near z0z_0, a polynomial looks like its Taylor expansion:

f(z0+ζ)=f(z0)+Akζk+(higher powers of ζ),f(z_0 + \zeta) = f(z_0) + A_k \zeta^k + (\text{higher powers of } \zeta),

where AkA_k is the first nonzero coefficient after the constant. For small ζ\zeta, the term AkζkA_k\zeta^k dominates the rest, and as ζ\zeta goes round a small circle, AkζkA_k\zeta^k goes kk times round a small circle. So the values f(z0+ζ)f(z_0 + \zeta) circle round f(z0)f(z_0), and some of them must lie on the side of f(z0)f(z_0) facing the origin. In the real landscape there are no “pits with a flat positive bottom”: every point not on a root has a downhill direction.

Explore Walking downhill on ∣f(z)∣|f(z)|

Try this. Drag the purple point z0z_0 anywhere in the left panel. The small circle round it is green where ∣f∣<∣f(z0)∣|f| < |f(z_0)| and orange where it is not, and the purple arrow points in Hardy's direction. Press “Step downhill” repeatedly. Then press “Go to a point where f′=0f' = 0” and look at both panels again.

f(z) = z3−3z+3z^3 - 3z + 3
z-plane: darker means smaller |f(z)|
xy
Z-plane near f(z₀): image of the small circle
XY
  • z₀ (drag), Hardy’s direction(s), and the image of the circle
  • part of the circle where |f| < |f(z₀)|
  • part where |f| ≥ |f(z₀)|
  • roots; on the right, the circle |Z| = |f(z₀)| and the direction of 0
z0=−1.6+1.5iz_0 = −1.6 + 1.5i, m=∣f(z0)∣=15m = |f(z_0)| = 15. The first nonzero coefficient in f(z0+ζ)=f(z0)+A1ζ+A2ζ2+⋯f(z_0 + \zeta) = f(z_0) + A_1\zeta + A_2\zeta^2 + \cdots is A1A_{1} (k=1k = 1, a=∣Ak∣=14.5a = |A_k| = 14.5). Hardy’s direction ϕ\phi with α+kϕ=μ+π\alpha + k\phi = \mu + \pi: 292°. On the circle of radius 0.25, the least value of ∣f∣|f| is 11.7, below m.
What to notice

Away from a root, once the radius is small enough the green arc is always there: about half the circle when k=1k = 1 (a large circle may have none). The right panel shows why: the image of the small circle is a loop round f(z0)f(z_0), and part of it crosses inside the dashed circle ∣Z∣=∣f(z0)∣|Z| = |f(z_0)|, closer to 00. Repeated steps run down into one of the dark wells, which are the roots. At the point where f′(1)=0f'(1) = 0 (so k=2k = 2) the image winds twice round f(z0)f(z_0) and there are two downhill arcs and two uphill ones, but still a way down.

The descent in the widget is not part of Hardy’s proof, which never moves anywhere: it only uses that a downhill direction exists at any point that is not a root. But stepping repeatedly in Hardy’s direction, with a step chosen so that ∣f∣|f| really decreases, is a version of Newton’s method, and it does find roots in practice.

Precise statements

Theorem Continuous functions on a square attain their bounds Hardy: Appendix II (B); cf. §103

Let F(x,y)F(x, y) be continuous on the closed square DD: ∣x∣≤R|x| \le R, ∣y∣≤R|y| \le R. Then FF is bounded on DD, and there are points of DD where it takes the values of its upper and lower bounds.

Hardy says this can be proved “much as in the last paragraph of §105”, that is, by repeated bisection, now of squares into four. (His first proof in §105 uses a Dedekind section, which has no analogue in two dimensions.) Continuity in two variables is defined in §108; Hardy does not prove the two-dimensional theorem in the main text.

Lemma d'Alembert's lemma Hardy: Appendix II (B)

Let ff be a polynomial and f(z0)≠0f(z_0) \ne 0. Then there is a point zz, as close to z0z_0 as we please, with ∣f(z)∣<∣f(z0)∣|f(z)| < |f(z_0)|.

Theorem The fundamental theorem of algebra Hardy: Appendix II (B)

Every polynomial of degree n≥1n \ge 1 with complex coefficients has a root. More precisely, ∣f(z)∣|f(z)| attains a least value mm over the whole plane, and m=0m = 0.

Proofs

Proof d'Alembert's lemma
Strategy

Expand ff round z0z_0 and keep only the first non-constant term AkζkA_k\zeta^k. Choose the direction of ζ\zeta so that AkζkA_k\zeta^k points from f(z0)f(z_0) straight towards the origin, and its length ρ\rho so small that the remaining terms cannot undo the gain.

Expand round z0z_0

Rearranging f(z0+ζ)f(z_0 + \zeta) in powers of ζ\zeta gives f(z0+ζ)=f(z0)+A1ζ+A2ζ2+⋯+Anζnf(z_0 + \zeta) = f(z_0) + A_1\zeta + A_2\zeta^2 + \dots + A_n\zeta^n, where the AA‘s do not depend on ζ\zeta and An=a0≠0A_n = a_0 \ne 0. Let AkA_k be the first nonzero one among A1,…,AnA_1, \dots, A_n.

Polar forms

Write f(z0)=meiμf(z_0) = m e^{i\mu} with m>0m > 0, Ak=aeiαA_k = a e^{i\alpha} with a>0a > 0, and ζ=ρeiϕ\zeta = \rho e^{i\phi}.

Make the tail small

Since ∣Ak+1ζk+1+⋯+Anζn∣≤ρk+1(∣Ak+1∣+⋯+∣An∣)|A_{k+1}\zeta^{k+1} + \dots + A_n\zeta^n| \le \rho^{k+1}(|A_{k+1}| + \dots + |A_n|) when ρ≤1\rho \le 1, we can choose ρ\rho so small that aρk<ma\rho^k < m and the tail gg satisfies ∣g∣<12aρk|g| < \tfrac12 a\rho^k. Then

f(z0+ζ)=meiμ+aρkei(α+kϕ)+g.f(z_0 + \zeta) = m e^{i\mu} + a\rho^k e^{i(\alpha + k\phi)} + g.

Point the main term at the origin

Choose ϕ\phi with α+kϕ=μ+π\alpha + k\phi = \mu + \pi. Then ei(α+kϕ)=−eiμe^{i(\alpha + k\phi)} = -e^{i\mu}, so the first two terms combine to (m−aρk)eiμ(m - a\rho^k)e^{i\mu}, a point on the segment from f(z0)f(z_0) towards 00.

Estimate

∣f(z0+ζ)∣≤(m−aρk)+∣g∣<m−12aρk<m.|f(z_0 + \zeta)| \le (m - a\rho^k) + |g| < m - \tfrac12 a\rho^k < m.

Since ρ\rho may be taken as small as we please, there are such points arbitrarily close to z0z_0.

Proof Every polynomial has a root (proof B)
Strategy

Restrict to a large square DD, on whose boundary ∣f∣|f| is larger than ∣f(0)∣|f(0)|. The least value of ∣f∣|f| on DD is attained at some z0z_0, which must be inside DD, and is then the least value over the whole plane. By d’Alembert’s lemma it cannot be positive.

∣f∣|f| is large far out

As in the previous lesson, ∣f(z)∣=∣a0∣∣z∣n∣1+η∣|f(z)| = |a_0||z|^n|1 + \eta| with ∣η∣→0|\eta| \to 0 as ∣z∣→∞|z| \to \infty. So we can choose RR such that ∣f(z)∣>∣f(0)∣|f(z)| > |f(0)| for every zz with ∣z∣≥R|z| \ge R, and in particular on and outside the boundary of the square DD of side 2R2R centred at 00.

A least value is attained in DD

F(x,y)=∣f(x+iy)∣F(x, y) = |f(x + iy)| is continuous and non-negative on DD, so it attains its lower bound mm at some point z0z_0 of DD.

z0z_0 is a minimum over the whole plane

m≤∣f(0)∣m \le |f(0)| because 00 is in DD. Points outside DD, and points on its boundary, have ∣f∣>∣f(0)∣≥m|f| > |f(0)| \ge m. So z0z_0 lies inside DD, and ∣f(z)∣≥m|f(z)| \ge m for every zz in the plane.

m=0m = 0

If m>0m > 0, d’Alembert’s lemma gives points zz arbitrarily close to z0z_0 with ∣f(z)∣<m|f(z)| < m, which contradicts the previous step. So m=0m = 0, that is, f(z0)=0f(z_0) = 0.

A hidden ingredient, and Littlewood’s repair

Look again at the step “choose ϕ\phi with α+kϕ=μ+π\alpha + k\phi = \mu + \pi”. To do it we need a number ζ\zeta with ζk\zeta^k in a prescribed direction; in other words we are solving an equation ζk=c\zeta^k = c. So proof B assumes the fundamental theorem for binomial equations zk=cz^k = c. We know these are solvable from de Moivre’s theorem (§48), but that rests on the circular functions. Is there a proof that avoids them?

Littlewood showed that there is, by running the same least-value argument on f(z)=zn−cf(z) = z^n - c itself.

  • Reduce to odd nn. Square roots can be found by pure algebra: if c=a+ibc = a + ib with b≠0b \ne 0 and r=a2+b2r = \sqrt{a^2 + b^2}, the roots of z2=cz^2 = c are ±(12(r+a)+i sgn⁡(b)12(r−a))\pm\left(\sqrt{\tfrac12(r + a)} + i\,\operatorname{sgn}(b)\sqrt{\tfrac12(r - a)}\right). If n=2vNn = 2^v N with NN odd, then vv successive square roots reduce zn=cz^n = c to wN=dw^N = d.
  • If the minimum point is not 00. For nn odd, let z0z_0 be a point where ∣zn−c∣|z^n - c| is least. If z0≠0z_0 \ne 0, then A1=nz0n−1≠0A_1 = n z_0^{n-1} \ne 0, so k=1k = 1, and choosing the direction only needs the solution of a linear equation.
  • If the minimum point is 00. Then f(ζ)=ζn−cf(\zeta) = \zeta^n - c. Taking ζ=±ρ\zeta = \pm\rho and ±iρ\pm i\rho, and using that nn is odd, ff takes the four values −c±ρn-c \pm \rho^n and −c±iρn-c \pm i\rho^n: small steps from −c-c in each of the four axis directions. One of these brings −c-c nearer to 00 (unless c=0c = 0), so 00 was not a minimum point after all.

Exercises

Exercise Taylor coefficients at a point ●○○

Let f(z)=z3−3z+3f(z) = z^3 - 3z + 3. Expand f(1+ζ)f(1 + \zeta) in powers of ζ\zeta. What are kk, aa, α\alpha and μ\mu at z0=1z_0 = 1, and which directions ϕ\phi does Hardy’s rule give? Check directly that ∣f(1+iρ)∣<1|f(1 + i\rho)| < 1 for small ρ>0\rho > 0.

Hint
(1+ζ)3=1+3ζ+3ζ2+ζ3(1 + \zeta)^3 = 1 + 3\zeta + 3\zeta^2 + \zeta^3.

Solution

f(1+ζ)=1+3ζ+3ζ2+ζ3−3−3ζ+3=1+3ζ2+ζ3f(1 + \zeta) = 1 + 3\zeta + 3\zeta^2 + \zeta^3 - 3 - 3\zeta + 3 = 1 + 3\zeta^2 + \zeta^3. So f(1)=1f(1) = 1 (m=1m = 1, μ=0\mu = 0), A1=0A_1 = 0 and A2=3A_2 = 3: k=2k = 2, a=3a = 3, α=0\alpha = 0. The rule α+2ϕ=μ+π\alpha + 2\phi = \mu + \pi gives ϕ=12π\phi = \tfrac12\pi (and ϕ=32π\phi = \tfrac32\pi, the other solution modulo 2π2\pi): straight up or down. Indeed f(1+iρ)=1−3ρ2−iρ3f(1 + i\rho) = 1 - 3\rho^2 - i\rho^3, and

∣f(1+iρ)∣2=(1−3ρ2)2+ρ6=1−6ρ2+9ρ4+ρ6<1|f(1 + i\rho)|^2 = (1 - 3\rho^2)^2 + \rho^6 = 1 - 6\rho^2 + 9\rho^4 + \rho^6 < 1

as soon as 9ρ2+ρ4<69\rho^2 + \rho^4 < 6, for instance for 0<ρ<120 < \rho < \tfrac12. Along the real axis, by contrast, f(1±ρ)=1+3ρ2±ρ3>1f(1 \pm \rho) = 1 + 3\rho^2 \pm \rho^3 > 1 for small ρ\rho: on the real line z0=1z_0 = 1 looks like a minimum.

Exercise Why the real line is not enough ●●○

The polynomial x2+1x^2 + 1 has no real root, and its least value on the real line, 11 at x=0x = 0, is positive. Locate exactly which step of d’Alembert’s lemma fails if we are only allowed to move along the real axis.

Solution

At x0=0x_0 = 0: f(0+ζ)=1+ζ2f(0 + \zeta) = 1 + \zeta^2, so k=2k = 2, A2=1A_2 = 1, α=0\alpha = 0, μ=0\mu = 0. Hardy’s rule asks for 2ϕ=π2\phi = \pi, that is ϕ=12π\phi = \tfrac12\pi: a vertical step, ζ=iρ\zeta = i\rho, giving f(iρ)=1−ρ2<1f(i\rho) = 1 - \rho^2 < 1. On the real axis only ϕ=0\phi = 0 and ϕ=π\phi = \pi are available, and both give ζ2=ρ2>0\zeta^2 = \rho^2 > 0, which increases ∣f∣|f|. The lemma needs every direction to be available, so that ζk\zeta^k can point anywhere: this is where the complex numbers are essential.

Exercise No hills either ●●○

Show that ∣f∣|f| has no local maximum: if ff is a non-constant polynomial and z0z_0 any point, there are points zz arbitrarily close to z0z_0 with ∣f(z)∣>∣f(z0)∣|f(z)| > |f(z_0)|.

Hint
Change one sign in the choice of ϕ\phi. Treat the case f(z0)=0f(z_0) = 0 separately.

Solution

If f(z0)=0f(z_0) = 0, any nearby point that is not a root will do, and the roots are finitely many. If f(z0)=meiμ≠0f(z_0) = m e^{i\mu} \ne 0, take kk, aa, α\alpha, ρ\rho and gg as in the lemma, but choose ϕ\phi with α+kϕ=μ\alpha + k\phi = \mu. Then f(z0+ζ)=(m+aρk)eiμ+gf(z_0 + \zeta) = (m + a\rho^k)e^{i\mu} + g, so ∣f(z0+ζ)∣≥m+aρk−∣g∣>m+12aρk>m|f(z_0 + \zeta)| \ge m + a\rho^k - |g| > m + \tfrac12 a\rho^k > m. (This is the polynomial case of the maximum modulus principle.)

Exercise A minimum on a closed set, but not on an open one ●○○

(a) Show that ∣ez∣|e^z| has no least value on the complex plane. (b) Which step of proof B fails for eze^z in place of a polynomial?

Solution

(a) ∣ex+iy∣=ex>0|e^{x + iy}| = e^x > 0 for every zz, and ex→0e^x \to 0 as x→−∞x \to -\infty, so the lower bound 00 is not attained.

(b) The first step: ∣ez∣|e^z| is not large far out in every direction (it is tiny for large negative xx), so we cannot confine the search for a minimum to a bounded square. d’Alembert’s lemma itself does hold for eze^z (it is never 00, and it has downhill directions everywhere); what fails is the existence of a minimum.

Exercise Square roots by algebra ●○○

Use the formula in Littlewood’s argument to find the square roots of 3+4i3 + 4i and of −5−12i-5 - 12i, and verify that the formula gives a square root of a+iba + ib in general.

Hint
For 3+4i3 + 4i: r=5r = 5.

Solution

3+4i3 + 4i: r=5r = 5, (5+3)/2=2\sqrt{(5 + 3)/2} = 2, (5−3)/2=1\sqrt{(5 - 3)/2} = 1, b>0b > 0: roots ±(2+i)\pm(2 + i). Check: (2+i)2=4+4i−1=3+4i(2 + i)^2 = 4 + 4i - 1 = 3 + 4i.

−5−12i-5 - 12i: r=13r = 13, (13−5)/2=2\sqrt{(13 - 5)/2} = 2, (13+5)/2=3\sqrt{(13 + 5)/2} = 3, b<0b < 0: roots ±(2−3i)\pm(2 - 3i). Check: (2−3i)2=4−12i−9=−5−12i(2 - 3i)^2 = 4 - 12i - 9 = -5 - 12i.

In general, with p=(r+a)/2p = \sqrt{(r + a)/2} and q=sgn⁡(b)(r−a)/2q = \operatorname{sgn}(b)\sqrt{(r - a)/2}: (p+iq)2=p2−q2+2ipq=a+isgn⁡(b)r2−a2=a+isgn⁡(b)∣b∣=a+ib(p + iq)^2 = p^2 - q^2 + 2ipq = a + i\operatorname{sgn}(b)\sqrt{r^2 - a^2} = a + i\operatorname{sgn}(b)|b| = a + ib.

Exercise Four small steps ●●○

Let w=−c≠0w = -c \ne 0 and t>0t > 0 be small. Show that at least one of w+tw + t, w−tw - t, w+itw + it, w−itw - it is nearer to 00 than ww is. (This is the last step of Littlewood’s argument.)

Hint
Write w=u+ivw = u + iv; at least one of u,vu, v is not 00.

Solution

Say u≠0u \ne 0 (otherwise swap the roles of the real and imaginary parts and use ±it\pm it). Take the sign s=−sgn⁡(u)s = -\operatorname{sgn}(u) and 0<t<2∣u∣0 < t < 2|u|. Then ∣w+st∣2=(u+st)2+v2=(∣u∣−t)2+v2<u2+v2=∣w∣2|w + st|^2 = (u + st)^2 + v^2 = (|u| - t)^2 + v^2 < u^2 + v^2 = |w|^2, since ∣∣u∣−t∣<∣u∣\bigl||u| - t\bigr| < |u|. Moving the real part towards 00 (by less than twice its size) shortens the vector.

Check yourself

Check yourself Enter a number

For f(z)=z3−3z+3f(z) = z^3 - 3z + 3 at z0=−1z_0 = -1 (where f′(−1)=0f'(-1) = 0), what is kk, the index of the first nonzero coefficient AkA_k with k≥1k \ge 1?

Why

f(−1+ζ)=5−3ζ2+ζ3f(-1 + \zeta) = 5 - 3\zeta^2 + \zeta^3: A1=f′(−1)=0A_1 = f'(-1) = 0 and A2=12f′′(−1)=−3≠0A_2 = \tfrac12 f''(-1) = -3 \ne 0. So k=2k = 2, and there are two downhill directions, now along the real axis (α=π\alpha = \pi, μ=0\mu = 0, so 2ϕ=02\phi = 0 or 2π2\pi).

Check yourself True or false?

Every continuous function F(x,y)≥0F(x, y) \ge 0 on the whole plane attains a least value.

Why

False: F=∣ex+iy∣=exF = |e^{x+iy}| = e^x has lower bound 00 and never reaches it. The theorem on bounds needs a closed bounded region. That is why proof B first shows that ∣f∣|f| is large outside a big square, so that the search for a minimum can be confined to the square.

Check yourself Choose one

At a point where f(z0)≠0f(z_0) \ne 0 and k=3k = 3, in how many of the directions round z0z_0 does ∣f∣|f| decrease to first order, and how are they arranged?

Why

To first order f(z0+ρeiϕ)≈f(z0)+aρ3ei(α+3ϕ)f(z_0 + \rho e^{i\phi}) \approx f(z_0) + a\rho^3 e^{i(\alpha + 3\phi)}, and the main term points “towards 00” (has a component opposite to f(z0)f(z_0)) for three arcs of ϕ\phi, each of width π/3\pi/3, alternating with three uphill arcs. Hardy’s three directions are the middles of the downhill arcs. In the widget, k=2k = 2 shows two and two.

Check yourself Choose one

Which fact does proof B, as Hardy first gives it, quietly assume?

Why

Choosing ϕ\phi with α+kϕ=μ+π\alpha + k\phi = \mu + \pi means finding ζ\zeta with ζk\zeta^k pointing in a given direction, i.e. solving ζk=c\zeta^k = c. De Moivre’s theorem provides this; Littlewood’s argument shows how to avoid it. No differentiability of ∣f∣|f| is used, and coefficients may be complex.

Hardy’s language

Summary

Summary
  • Far out ∣f(z)∣|f(z)| is large, so ∣f∣|f| attains a least value mm at some point z0z_0 of a big square, and this is its least value over the whole plane.
  • d’Alembert’s lemma: if f(z0)=meiμ≠0f(z_0) = me^{i\mu} \ne 0 and Ak=aeiαA_k = ae^{i\alpha} is the first nonzero Taylor coefficient, a small step ρeiϕ\rho e^{i\phi} with α+kϕ=μ+π\alpha + k\phi = \mu + \pi gives ∣f∣<m−12aρk|f| < m - \tfrac12 a\rho^k.
  • So m=0m = 0: the least value is a root.
  • The step uses complex directions (it fails on the real line) and assumes ζk=c\zeta^k = c is solvable; Littlewood’s variant removes that assumption using square roots and four axis-directions.