The optimist's question after two guides of continuation
By now analytic continuation feels almost greedy. You start with a power series in a disk, find it agrees with a function on an overlap, and rebuild that function on a fresh patch — then another, then another, crawling across the plane. The first guide gave you the engine and the uniqueness of continuation that makes every step forced; the second taught you to drag a function element along a path and warned that a loop can hand back a different branch. A natural optimism follows: surely, given enough patience, you can continue anything to almost everywhere. This guide is where that optimism meets an honest wall.
First recall what does NOT stop continuation. A lone singularity — a pole, a branch point, an essential singularity — is never a barrier; it is a pebble you walk around. If a power series has disk of convergence of radius 1 because there is exactly one bad point sitting on the circle, you simply re-expand the function about a center off to the side, get a new disk that pokes past the old circle on the far side of the pebble, and keep going. The radius of a Taylor series only ever reaches the nearest singularity — it tells you where THIS disk ends, not where the function ends.
So the question sharpens. A single singularity on the boundary circle is escapable, and so is any finite handful — there is always clear water between them to slip a new disk through. What if the singularities are not finite, though? What if they pile up along the whole boundary circle so thickly that no matter where you aim a new disk, its arc immediately runs into another bad point? Then there is no gap to slip through, nowhere to plant a foot on the other side. That dense, unbroken wall of singularities is a natural boundary, and crossing it is not hard but impossible.
Building a wall by hand: the gap series
Abstract talk of 'densely packed singularities' is unconvincing until you meet one, so here is the cleanest example in all of complex analysis. Consider the series f(z) = sum z^(2^n) over n = 0, 1, 2, ... — that is, z + z^2 + z^4 + z^8 + z^16 + ... where only the powers that are powers of two appear. This is a lacunary series, from the Latin lacuna, a gap: between each term and the next there is a yawning gap of skipped powers, and that gap is exactly what builds the wall.
Its radius of convergence is exactly 1: for |z| < 1 each term z^(2^n) is tiny and the series converges to a perfectly nice holomorphic function inside the unit disk, while for |z| > 1 the terms blow up. So f starts life as an honest holomorphic function on the open unit disk. The whole drama is about the boundary, the unit circle |z| = 1. The claim — which feels outrageous the first time you hear it — is that EVERY point of that circle is a singularity, and so f cannot be continued one millimeter past it anywhere. The unit circle is f's natural boundary.
The mechanism is a beautiful piece of bookkeeping. Look at what f does as z runs out to the boundary along a ray pointing at a root of unity. Take z = r times a 2^k-th root of unity and push r up toward 1. For every n at least k, the angle 2^n times the root's angle is a whole multiple of 2 pi, so e^(i times that angle) = 1, and those infinitely many terms become a real series of positive numbers z^(2^n) that marches off to infinity as r approaches 1. Only finitely many early terms stay finite. So f blows up along every such ray.
f(z) = z + z^2 + z^4 + z^8 + z^16 + ... (powers 2^n)
converges and is holomorphic for |z| < 1
diverges for |z| > 1
At z = r * w, where w is any 2^k-th root of unity:
for n >= k: (z)^(2^n) = r^(2^n) * w^(2^n)
= r^(2^n) * 1 (angle is a multiple of 2 pi)
so the tail r^(2) + r^(4) + r^(8) + ... -> +infinity as r -> 1
The 2^k-th roots of unity, over ALL k, are DENSE on the circle
=> f blows up on a dense set => no point of |z|=1 is regular
=> the unit circle is a NATURAL BOUNDARYWhy a dense set of bad points seals every gap
The key word is dense. The 2^k-th roots of unity, collected over all k = 1, 2, 3, ..., are sprinkled around the circle with no gaps: between any two points of the circle, however close, there is one of these roots. Since f blows up at each of them, the singular points of f form a dense subset of the unit circle. Now try to escape. To continue past the boundary you would need an arc of the circle that is free of singularities — clear water to float a new disk through. But a dense set leaves no arc free, however short. Every door you reach for is already nailed shut by a singularity sitting on it.
Compare this with the friendly case from the first section to feel the difference. The function 1/(1 - z) also has radius of convergence 1, also from a singularity on the unit circle — but only ONE, at z = 1. The rest of the circle is clear, so you re-center anywhere away from z = 1, your new disk reaches across the circle through the clear water, and the function continues happily to the whole plane minus the single point z = 1. The lacunary f has no clear water anywhere. That is the entire distinction between a removable obstacle and a true boundary: one singularity versus a dense crowd of them.
How to smell a natural boundary coming: the gaps
Was the choice of powers 2^n special, or is something general going on? Something general. The exponents 1, 2, 4, 8, 16, ... grow so fast that the GAPS between consecutive ones grow without bound — each exponent is at least double the one before. A foundational result (Hadamard's gap theorem, the headline you will meet by name later) says: if the non-zero terms of a power series of radius 1 sit at exponents whose gaps grow at least like a fixed ratio bigger than 1, then the circle of convergence is automatically a natural boundary. The coefficients can be anything; it is the SPARSENESS of the powers, the silence between the surviving terms, that locks the function inside its disk.
There is a vivid way to feel why gaps do this. A nice continuable function, expanded as a power series, has coefficients that 'know' about a singularity on the boundary in a coordinated way — they conspire to make the function blow up at one special spot and behave everywhere else. A gap series refuses to coordinate: with whole blocks of coefficients forced to zero, the surviving terms cannot single out any one boundary direction as special, so they treat all directions alike — and 'singular in the same way in every direction' is exactly what a natural boundary is. Symmetry forbids any escape hatch because no point on the circle is allowed to be less singular than its neighbors.
What a natural boundary really means
Step back and ask what we have actually learned about a function with a natural boundary. The unit disk is not just the first place we happened to define f — it is f's entire universe, the largest domain on which f can ever exist as a holomorphic function. There is no bigger function out there of which our f is a fragment. In the language of the previous guides, the complete analytic function you would build by exhausting every possible continuation is just f on the disk and nothing more. The continuation process, run to its absolute limit, returns the disk you started with.
Hold this honestly alongside the uniqueness theorem so the two do not seem to clash. Uniqueness says: IF a continuation exists across some region, it is forced and unique — there is never a choice about HOW to extend. A natural boundary says: sometimes a continuation does not exist at all — there is nowhere to extend TO. The two are perfectly consistent. Uniqueness is a statement about the impossibility of ambiguity; a natural boundary is a statement about the impossibility of extension. Neither ever promised you that every function reaches the whole plane; that was the optimist's hope, and this guide is its correction.
This also resets your map of the territory for the rest of the rung. Functions come in genuinely different kinds. Some, like 1/(1 - z), continue almost everywhere, blocked only at isolated points you route around. Some, like the square root and the logarithm coming up next, continue everywhere but come back as a DIFFERENT branch after a loop — the multivalued case the monodromy theorem of guide two governed, and the case Riemann surfaces will tame. And some, like our gap series, simply stop dead at a wall and have no life beyond it. Knowing which kind you are holding — escapable obstacle, multivalued ambiguity, or impassable wall — is half the art of this subject.