Asking whether a real function is differentiable is asking whether the difference quotient converges as $h\to 0$ — and $h$ can only arrive from the left or the right. A complex $h$ can approach 0 from infinitely many directions in the plane, and complex differentiability demands that all of those paths return the same limit. That is not one condition; it is infinitely many conditions imposed at once.
What you buy with that price is rigidity. A real smooth function is putty: you can build one that vanishes identically on $[-1,1]$ yet is nonzero outside, infinitely differentiable throughout — squeeze it locally and nothing elsewhere notices. A holomorphic function is crystal: know it on any small disc and its values on the entire connected region are already pinned down. One complex derivative buys you infinitely many, plus equality with your own Taylor series everywhere.
Write $f(z)=u(x,y)+i\,v(x,y)$. Complex differentiability is equivalent to the Cauchy–Riemann equations:
Here $u$ and $v$ are the real and imaginary parts. Together the two equations say: differentiating along the real axis and along the imaginary axis must return the same complex number. Geometrically that is equivalent to the gradients of $u$ and $v$ being orthogonal and of equal length — so the derivative matrix of $f$ can only be "rotation × scaling", with no stretching and no shear. Differentiate each equation once more and subtract, and out drops $\nabla^2u=\nabla^2v=0$: the real and imaginary parts of a holomorphic function are automatically harmonic.
A purely local differential constraint forces global conclusions. Liouville's theorem: a function holomorphic and bounded on the whole plane must be constant — from which the fundamental theorem of algebra takes three lines (if a polynomial $p$ had no root, $1/p$ would be a bounded entire function, so $p$ is constant). The maximum modulus principle: $|f|$ attains its maximum on the boundary; no peaks are permitted in the interior. All the information lives on the boundary — a theme that recurs later in the Dirichlet problem and even in the holographic principle.
Two-dimensional electrostatics and ideal fluid flow: let $u$ be the potential and $v$ the stream function, and Cauchy–Riemann automatically enforces irrotational, source-free flow with equipotentials orthogonal to streamlines everywhere. In signals and materials, causality (no response before its stimulus) is equivalent to the transfer function being holomorphic in the upper half-plane, which yields the Kramers–Kronig relations: a medium's refractive index and its absorption spectrum determine each other, so measuring one lets you compute the other — an experimental dividend paid directly by holomorphy.
Picture $f$ as a flow field on the plane. Holomorphic means source-free and irrotational, so the circulation around any closed curve is zero (Cauchy's theorem). Singularities are the vortices pinned into the plane — the pinholes where holomorphy fails.
Contour integration therefore becomes something close to absurd: it records only which vortices you enclosed and how many times you circled each. How long the curve is, how badly it meanders, how far it passes from a singularity — all irrelevant. Shrink it, knead it, drag it about; as long as you never cross a singularity, the value does not shift by a single digit.
$\gamma$ is a closed curve, the $z_k$ are the enclosed isolated singularities, $n(\gamma,z_k)$ is the winding number (how many times the curve circles that point, counterclockwise positive), and $\mathrm{Res}(f,z_k)$ is the residue — the coefficient of $(z-z_k)^{-1}$ in the Laurent expansion of $f$ near $z_k$.
Why exactly the power $-1$? Going once around a small circle, $\oint(z-z_0)^n dz$ vanishes for every $n\neq-1$ — the antiderivative is single-valued, so the trip cancels itself out. Only for $n=-1$ is the antiderivative $\log$, and $\log$ does not return to its original value after a loop; it comes back exactly $2\pi i$ richer. The whole theorem hangs off the multivaluedness of the logarithm.
A continuous integral along a curve collapses into a handful of algebraic coefficients at isolated points. Local data completely determines a global integral — the earliest specimen of "the whole equals the sum of the local", replayed afterwards in Stokes' theorem, de Rham cohomology, and the index theorems. It also carries a trespasser's thrill: integrals that refuse to yield on the real line land immediately once lifted into the complex plane and looped, such as $\int_{-\infty}^{\infty}\frac{\sin x}{x}\,dx=\pi$.
Every propagator in quantum field theory is a contour integral, and whether you route above or below a pole (the $i\epsilon$ prescription) decides whether you obtain the retarded or the advanced Green's function — causality encoded as a choice of detour. The Nyquist criterion in control theory: closed-loop stability equals the number of times the open-loop frequency response winds around $-1$ (the argument principle, a corollary of the residue theorem). In algorithm analysis, generating-function coefficients are $a_n=\frac{1}{2\pi i}\oint\frac{F(z)}{z^{n+1}}dz$, and saddle-point methods turn that into asymptotics for combinatorial counts.
Since a holomorphic map acts locally as nothing but rotation and scaling, the angle at which two curves cross is preserved untouched. At large scale the region may be warped beyond recognition; at infinitesimal scale every intersection keeps its angle to the last decimal.
Picture an infinitely stretchable rubber map: pull it, bend it, do as you please — only shear is forbidden. The restriction sounds mild and turns out to be brutal: it makes almost every shape into the same shape.
Let $\Omega\subsetneq\mathbb{C}$ be a simply connected open region ("simply connected" = no holes; "$\subsetneq$" = not the whole plane). Then there exists a biholomorphic map
"Biholomorphic" means $f$ is invertible with both $f$ and $f^{-1}$ holomorphic. Once you specify an interior point that maps to 0 and fix the argument of the derivative there, $f$ is unique.
A heart, a star, an airfoil cross-section, an endlessly meandering river channel — provided there are no holes, complex analysis sees one and the same object: the unit disc. An apparently inexhaustible space of shapes is crushed down to a point. The converse is just as sharp: put a hole in the region and invariants appear at once (the ratio of an annulus's inner and outer radii cannot be changed conformally). Conformal mapping draws the cleanest line yet between "classification is trivial" and "continuous parameters appear".
The Joukowski transform carries a circle to an airfoil cross-section, so the solution for flow past a cylinder transfers straight across and the lift can be read off — the workhorse of early aerodynamics. Effective resistance in chip routing, groundwater seepage, and electrostatic fields with awkward electrode shapes are all solved by conformally pulling a nasty boundary into a rectangle. The Mercator chart is conformal, which is why rhumb lines are straight, at the cost of area distortion (Greenland looks like Africa). Flattening a 3D surface for texture mapping in computer graphics is discrete conformal geometry.
$\sqrt{z}$ has two values and $\log z$ has infinitely many. The textbook remedy is to "take the principal value": slit the plane along the negative real axis and force the function to be single-valued. But that cut is an artifact — where you place it is pure convention, and the function itself has no trouble there at all.
Riemann inverted the question: do not repair the function, repair its domain. Take two copies of the plane, glue them crosswise along the slit, and build a two-storey spiral staircase; on this new surface $\sqrt{z}$ is perfectly single-valued and analytic. Walk once around the origin and you have not returned to your starting point — you have climbed a floor. Only after two loops does the path close. Multivaluedness has not been abolished; it has been translated into topology.
A Riemann surface is a one-dimensional complex manifold: a connected Hausdorff space covered by a family of coordinate charts whose transition maps are holomorphic wherever two charts overlap. "One complex dimension" means it looks locally like a patch of $\mathbb{C}$, so as a real space it is two-dimensional — a surface.
How many branches the function has ↔ how many sheets the surface has; branch points ↔ the seams where sheets are stitched. The infinitely many values of $\log$ are simply an infinite spiral. More striking still is the uniformization theorem: every simply connected Riemann surface is one of exactly three things — the sphere, the plane, or the unit disc, corresponding to constant positive, zero, and negative curvature. And compact Riemann surfaces are the same objects as algebraic curves, with the genus $g$ serving at once as a topological, an algebraic, and an analytic quantity, threaded into a single equation by the Riemann–Roch theorem. Here geometry, algebra, and analysis are not "related" — they are three notations for one thing.
A string's worldsheet is a Riemann surface, and the loop order of the perturbative expansion is its genus. A genus-1 Riemann surface is a torus — that is, an elliptic curve, the shared core object of ECC public-key cryptography and of Wiles' proof of Fermat's Last Theorem; one geometric body holding up both information security and modern number theory. In engineering, where you place the branch cuts of a multivalued transfer function directly determines the stability region of a filter.