Why a bound is worth more than a value
In the previous guide you learned to compute a contour integral by hand: pick a parametrization z(t), feed it in, and grind out integral of f(z(t)) z'(t) dt. That is honest work, and for a circle or a line segment it goes through cleanly. But here is a secret of the whole subject — most of the time you do not want the value of an integral at all. You want to prove it is small, often that it shrinks to zero as some contour grows or some radius shrinks. A single sharp estimate can replace a page of exact computation that you would only throw away.
Think of where this bites. To prove Cauchy's theorem you will cut a region into ever-smaller triangles and show the leftover error vanishes. To prove that an integrand has no pole hiding inside, you let a big circle blow up and watch the integral die. To prove Liouville's theorem — that a bounded entire function must be constant — you bound an integral over a huge circle and let the radius run to infinity. Every one of these arguments needs the same gadget: a way to say "this integral is at most such-and-such, and such-and-such goes to zero." That gadget is the ML inequality.
Stating the inequality, and what M and L mean
Let f be continuous on a contour C, let L be the length of C, and let M be any bound on the size of f along C — that is, a number with |f(z)| at most M for every z on the curve. Then the modulus of the integral is at most M times L. The name writes itself: M for the modulus bound, L for the length. It is the cleanest tool in the toolbox, and it almost never fails to help.
| integral over C of f(z) dz | <= M * L
where L = length of C = integral over C of |dz|
M = any number with |f(z)| <= M for all z on CTwo cautions before you reach for it. First, M only has to be an upper bound, not the exact maximum — a generous, easy-to-find M is perfectly legal and usually all you need; tightening it is wasted effort unless the slack actually breaks your argument. Second, the inequality bounds the modulus of the integral; it tells you nothing about the integral's argument or its sign, and it can never prove an integral is large. It is a one-directional tool, and that one direction is exactly the one the theorems ahead ask for.
Where it comes from: the triangle inequality, integrated
The proof is short and worth carrying in your head, because it shows the inequality is not a deep new fact but an old one — the triangle inequality — applied to a sum that has become an integral. Recall that a contour integral is the limit of sums f(z_k) times (small step delta z_k). The triangle inequality says the modulus of a sum is at most the sum of the moduli, so the modulus of the whole approximating sum is at most the sum of |f(z_k)| times |delta z_k|.
Now bound each |f(z_k)| by M and pull it out, leaving M times the sum of the step lengths |delta z_k|. But that remaining sum of tiny step-lengths is precisely the polygonal approximation to the arc length of C, which converges to L. Pass to the limit and you have it: |integral| at most M times L. In a single line, taking the modulus inside the integral can only make things bigger, |f(z)| is at most M, and the integral of |dz| is the length L.
A worked estimate: the integral over a big circle
Take f(z) = 1 / (z^2 + 1) and let C_R be the circle |z| = R, traversed once, for some large R. We will not compute the integral; we will only show it shrinks as R grows. The walk-through below is the pattern you will repeat in dozens of later problems, so it is worth doing slowly once.
- Find L. The contour is a full circle of radius R, so its length is L = 2 pi R. (Half this if you only use the upper semicircle — a case that comes up constantly later.)
- Bound |f| from above by making the denominator small from below. On C_R we have |z| = R, and the reverse triangle inequality gives |z^2 + 1| at least |z^2| - 1 = R^2 - 1. So |f(z)| = 1 / |z^2 + 1| is at most 1 / (R^2 - 1). Take M = 1 / (R^2 - 1).
- Multiply: |integral| is at most M * L = 2 pi R / (R^2 - 1).
- Read off the limit. As R goes to infinity, 2 pi R / (R^2 - 1) behaves like 2 pi / R and goes to zero. So the integral of 1/(z^2+1) over the big circle vanishes in the limit — no computation of the value ever needed.
Notice the moving part that made it work: the denominator grew like R^2 while the length only grew like R, so M shrank faster than L grew, and the product went to zero. That race — *does the integrand decay faster than the path lengthens?* — is the question behind every vanishing-arc argument, including the ones you will meet when the residue theorem is used to crack real integrals. The reverse triangle inequality, |a + b| at least the absolute difference of |a| and |b|, is the everyday workhorse for finding M; keep it close.
Sharp enough, and where it is too blunt
How good is the bound? Sometimes it is exactly tight. Integrate f(z) = 1 over the unit circle: M = 1, L = 2 pi, so ML says the modulus is at most 2 pi — and the true value, integral of dz around the circle, is 0, comfortably under the bound but reminding you that ML caps the size, not the value. A better demonstration of tightness: integral of (1/z) dz around the unit circle equals 2 pi i, whose modulus is 2 pi; here |1/z| = 1 on the circle, so M = 1 and L = 2 pi give the bound 2 pi exactly. The estimate is achieved. That equality case happens precisely when |f| is constant on C and f(z) dz keeps a constant argument all the way round.
And where is it blunt? When the integrand is large on a tiny stretch of C and small everywhere else, a single flat M — forced to cover the worst point — wildly overcounts the long quiet stretch. The honest fix is exactly what the proof suggested: split C into sub-arcs and apply ML separately on each, with a smaller M where f is small. The total bound is the sum, and it can be enormously sharper than one M over the whole curve. ML is a starting estimate you refine, not a final answer you accept.