Three behaviours, and why the third is different in kind
By now the three-way classification of an isolated singularity is familiar: read the principal part of the Laurent series and count its negative powers. None means removable, finitely many means a pole, infinitely many means essential. The first two guides in this rung walked you through telling them apart and measuring the order. This last guide is about what that third box really means — because 'infinitely many negative powers' sounds like a bookkeeping detail, and it is anything but.
There is a clean way to see the difference through limits alone, the limit test. At a removable singularity the function stays bounded and approaches a finite limit; you fill the hole and walk away. At a pole the modulus |f(z)| marches off to infinity no matter how you approach z_0 — it tends to a single, definite limit, the point at infinity. The essential case is exactly the leftover: the limit of f(z) as z approaches z_0 simply does not exist, not even as infinity. The function refuses to settle.
Meet the wild one: e^(1/z) at the origin
The cleanest essential singularity is f(z) = e^(1/z) at z_0 = 0. Take the ordinary exponential series e^w = 1 + w + w^2/2! + w^3/3! + ... and substitute w = 1/z. You get a Laurent series whose principal part never ends: e^(1/z) = 1 + 1/z + 1/(2! z^2) + 1/(3! z^3) + ... . Infinitely many negative powers, all nonzero — by the count, this is an essential singularity, textbook case zero.
Now watch it refuse to settle. Approach 0 along the positive real axis, z = x with x going to 0 from above: then 1/z = 1/x goes to +infinity, so e^(1/z) blows up to +infinity. Approach instead along the negative real axis, z = -x: then 1/z = -1/x goes to -infinity, so e^(1/z) tends to 0. Two different directions, two utterly different destinations — one infinite, one zero. That alone kills any hope of a single limit, so 0 cannot be a removable singularity (no finite limit) and cannot be a pole (|f| would have to go to infinity from every direction, but it goes to 0 along one of them).
But the real shock is not that the limit fails to exist — it is how thoroughly it fails. Pick any nonzero target value, say w = 3 + 4i. Can we make e^(1/z) hit it with z as close to 0 as we please? We need 1/z = log(3 + 4i), and the logarithm is multivalued: log(3 + 4i) + 2 pi i k works for every integer k. As k grows, those values march off to infinity, so 1/z is huge, so z = 1/(log(3+4i) + 2 pi i k) is tiny — arbitrarily close to 0. The function hits 3 + 4i infinitely often in every neighbourhood of the singularity. And nothing was special about 3 + 4i.
Casorati-Weierstrass: the image is everywhere
What e^(1/z) does by hand is a theorem in general. The Casorati-Weierstrass theorem says: if z_0 is an essential singularity of f, then on every punctured disk 0 < |z - z_0| < r — no matter how small you shrink r — the image f(z) is dense in the whole complex plane. Dense means: for any target value w and any tolerance epsilon, you can find a z that close to z_0 with |f(z) - w| < epsilon. The function comes within a hair of every value, again and again, in any tiny ring hugging the singularity.
The proof is a lovely short argument by contradiction that leans entirely on the removable and pole alternatives you already understand. Suppose the image were not dense. Then some disk around some value w is missed entirely: |f(z) - w| stays at least delta away from 0 on the punctured neighbourhood. Form g(z) = 1 / (f(z) - w). Because f - w never gets near 0, g is bounded near z_0, so by the Riemann theorem g has a removable singularity there — g extends holomorphically. Then f = w + 1/g, and reading off the singularity of f from g forces it to be removable or a pole, never essential. That contradicts our assumption, so the image must be dense after all.
Picard: dense is not even the whole truth
Casorati-Weierstrass says the function comes arbitrarily close to every value. The astonishing Great Picard theorem says it does far more: near an essential singularity, f actually attains every complex value — infinitely often — with at most a single exception. Not close to; equal to. In every punctured neighbourhood, no matter how small, the equation f(z) = w has infinitely many solutions for every w except possibly one exceptional value.
Our example shows the exception is real and unavoidable: e^(1/z) is never zero anywhere (the exponential never vanishes), so w = 0 is the one value it genuinely skips. Every other complex number it hits infinitely often in any disk around 0, exactly as we found by hand for 3 + 4i. One missed value, all the rest attained infinitely many times — that is the precise face of the wildness. There is also a Little Picard theorem, its global cousin: a non-constant entire function omits at most one complex value (e^z again omits only 0). Little Picard is what you get by applying the great theorem to the essential singularity that every non-polynomial entire function hides at infinity.
near an essential singularity z_0:
Casorati-Weierstrass : image is DENSE in C (comes arbitrarily close to every w)
Great Picard : ATTAINS every w in C, (equals every w, infinitely often,
except at most one value with at most one exception)
example e^(1/z) : the skipped value is w = 0 (e^w is never 0)Where this leaves you, and where it points
Step back and the rung holds together. A Laurent series gives you the principal part; counting its negative powers sorts every isolated singularity into removable, pole, or essential; the order of zeros and poles measures the tame cases; meromorphic functions are the well-behaved world where only poles are allowed. Essential singularities are exactly what meromorphic functions forbid — and now you know why they are forbidden: a single essential point makes the function visit (nearly) the whole plane in any ring around it, which is incompatible with the orderly value-at-infinity behaviour a meromorphic function demands.
There is one loose thread worth naming: we kept saying isolated singularity, a bad point with a clean ring of good behaviour around it. Not every singularity is isolated — Log z along its branch cut, or the point 0 for sin(1/z) (which has a singularity that is a limit of other singularities), behave differently and fall outside this whole classification. The Laurent-series machinery is precisely a tool for the isolated case; honoring that boundary is part of using it correctly.
Looking forward, the single coefficient a_{-1} of any Laurent series — the principal part's leading term — is about to take centre stage as the residue, and the next rung turns it into a machine for evaluating real integrals you could never crack by real-variable methods. The wildness you met here also feeds into deeper value-distribution theory (Picard is its doorway). But for this rung the picture is complete: removable, pole, essential — three genuinely different kinds, and you now understand the third not as a label but as a riot of values crowded into an arbitrarily small ring.