Bounded, Therefore Constant

In 1844 Cauchy proved that a bounded entire function is constant. Liouville published the same statement about elliptic functions three years later, and the theorem took Liouville’s name.

The naming is the least useful thing about it. The conditions are where the work happens.

Entire means holomorphic on the whole complex plane. Bounded means one finite ceiling applies everywhere, not merely across the part somebody happened to inspect. Given both, Cauchy’s estimate forces the first derivative to vanish at every point, and there is no room left for a non-constant extension.

That is what a continuation is when it has earned the word. Not a curve carried past its last plotted point. The only member of a class still compatible with what has already been fixed.

One derivative fixes all the others #

Cauchy’s differentiation formula supplies every derivative at once, from a single contour integral. A holomorphic function is therefore infinitely differentiable and analytic: on every disk inside its domain it equals its own Taylor series. The Cauchy estimates make the constraint quantitative, bounding each derivative in terms of the function’s size on a surrounding circle and the radius of that circle.

This is not the ordinary relationship between a sampled series and a fitted line.

It is a class membership test with consequences attached. Complex differentiability at every point of an open domain has already excluded almost all possible local behaviour before anyone asks what the function does.

Real functions have no such guarantee. A real function can be infinitely smooth at a point and still fail to equal its Taylor series anywhere near it, and the standard counterexample has been in textbooks for a century. Holomorphicity closes that gap. Nothing in a data series does.

Agreement near a point can settle a whole domain #

The identity theorem gives the operational form. Two functions holomorphic on the same connected domain, agreeing on a set with an accumulation point inside it, agree everywhere on that domain.

A sequence of distinct agreement points converging inside the domain is enough. An open sub-region is far more than enough.

Analytic continuation runs on that fact, joining overlapping power-series expansions. Once two patches agree on an accumulating set they stop being two candidates. They are one function on the overlap, and where the continuation succeeds, one answer on the enlarged domain.

The domain is doing visible work in that sentence. The theorem does not say an agreement set authorises a value wherever somebody would like one. It says what follows on a connected domain under holomorphicity.

Remove those words and the conclusion stops being a theorem and becomes a preference.

The failures are part of the function #

Reference works do not record a formula and leave the rest implied. They record admissible inputs, outputs and failures, in the same entry, at the same level of prominence.

Arcsin has domain from minus one to one because sine was restricted before being inverted. The exponential has the whole plane for a domain and everything except zero for a range. The Gamma function excludes every non-positive integer, and its poles there are simple, with known residues.

The zeta function makes the distinction between an expression and the object it defines. Its series converges only to the right of one. Its analytic continuation is defined everywhere except a single simple pole. The continued function is not the series asserted further than it converges; it is a constrained object reached from the series, and the constraint is what makes it unique.

Singularities are therefore not caveats added after a result. They state which extensions remain legal.

The most compact failure is the logarithm. Before a branch cut is imposed it is not a function at all but a relation, since every nonzero input has infinitely many logarithm values. The square root has two. The principal branch that everyone uses is a choice among valid determinations, and the symbols do not carry it unaided.

An expression becomes usable as a function by carrying its domain and its determination along with it. Nothing else supplies them.

There is no identity theorem for a fitted sample #

Two valuation models can agree on every observed value and diverge the moment they reach an unobserved interval.

That divergence is not numerical noise around one true continuation. It is evidence that the continuation came out of a chosen structural class — a growth rule, a singularity assumption, a branch convention, an assumption that no break occurs — and that the two models chose differently.

There is a nearby problem worth keeping separate. Interpolation inside densely observed data is also model-dependent, and it can be tested, because a local holdout is available. Beyond the last observation the entire target region is the holdout.

A curve can pass every in-sample check ever devised and still have no test at all of the part it is being used to supply.

And complex analysis has a constraint no financial series possesses. A holomorphic function cannot change regime halfway across a connected domain and remain the same holomorphic function. That is not a convention; it is the content of the theorem.

A market changes regime whenever it likes.

A power series continues to exactly one function, and not because the series is well behaved. Because somebody named the class it belonged to before extending it.

A fitted curve names no class. It runs past the last observation on the strength of a shape, and the values it produces out there were manufactured by the assumption rather than measured. Which is perfectly usable, as long as the assumption is written on the chart next to them.