Why circles go to circles
The cleanest argument is to break f into the three pieces above and
note each one preserves the class of circlines.
Step 1 — divide out (the case c \ne 0). Polynomial
division on \dfrac{az + b}{cz + d} peels off a constant:
\frac{az + b}{cz + d} = \frac{a}{c} + \frac{b - \frac{ad}{c}}{cz + d} = \frac{a}{c} - \frac{ad - bc}{c\,(cz + d)}.
Step 2 — read off the chain of simple maps. Building
z up to that expression uses only the three primitives, in order:
z \;\xrightarrow{\ \times c,\ +d\ }\; cz + d \;\xrightarrow{\ 1/(\cdot)\ }\; \frac{1}{cz + d} \;\xrightarrow{\ \times(-\frac{ad-bc}{c}),\ +\frac{a}{c}\ }\; f(z).
Step 3 — each primitive preserves circlines. Translations and
scaling–rotations obviously send circles to circles and lines to lines. The one substantive
fact is that inversion z \mapsto 1/z maps circlines to
circlines (a line not through the origin becomes a circle through the origin, and so
on). Composing maps that all preserve circlines gives a map that preserves circlines —
done.
The model example: half-plane to disk
The single most useful Möbius map in analysis is
f(z) = \frac{z - i}{z + i}.
It sends the upper half-plane
\{\operatorname{Im} z > 0\} bijectively onto the open unit
disk \{|w| < 1\}, carrying the boundary real axis onto the
boundary unit circle. Track three boundary points and the circline rule does the rest.
Step 1 — the point 0.
f(0) = \frac{0 - i}{0 + i} = \frac{-i}{i} = -1.
Step 2 — the point 1.
f(1) = \frac{1 - i}{1 + i} = \frac{(1 - i)^2}{(1 + i)(1 - i)} = \frac{-2i}{2} = -i.
Step 3 — the point \infty. With
a = b = 1, c = 1,
d = i,
f(\infty) = \frac{a}{c} = \frac{1}{1} = 1.
Three points of the real axis — 0, 1, \infty — land on
-1, -i, 1, three points of the unit circle. Since a circline is
pinned down by three points, the whole real axis must map to the whole unit circle, and the
half-plane on one side to the disk on the inside.
"Circles-and-lines to circles-and-lines" does not mean lines stay lines and
circles stay circles. A perfectly straight line can map to a circle, and a
circle can map to a straight line — that is exactly what the model example above does, sending
the real axis (a line) onto the unit circle. The resolution is to stop treating lines and
circles as separate species: in the extended plane a line is just a circle that passes
through the point at infinity \infty. What a Möbius map
preserves is the single family of generalized circles ("circlines") — circles
or lines — not lines and circles separately. So the trap is expecting
\operatorname{Im} z = t to come out straight; whenever the image
circline misses \infty, your line has curled up into an honest
circle.
The leap to the Riemann sphere is what makes Möbius maps so tidy. On the ordinary plane the
point z = -d/c looks like a catastrophe — the denominator
vanishes. Add the single point \infty and the catastrophe becomes
a smooth bijection: -d/c simply goes to
\infty, and \infty goes to
a/c. Lines and circles stop being different species — both are
just circles on the sphere, one of them passing through the north pole. The algebra of these
maps is the algebra of invertible 2 \times 2 matrices
\begin{pmatrix} a & b \\ c & d \end{pmatrix}, which is why
composing two Möbius maps is the same as multiplying their matrices.