Singularities and Classification

Suppose a function is holomorphic everywhere on a punctured disk around a point. Its behavior at that point, though, is unknown or undefined. What can happen there? How much can we learn about the missing value from the function's local expansion? This question sits at the heart of complex analysis. Isolated singularities are the loci where the smooth machinery of holomorphic functions threatens to break down. Yet it breaks down in a shockingly rigid way. There are exactly three qualitatively different behaviors. Each is completely diagnosed by a single piece of algebraic data. That data comes from the Laurent Series of the function. This trichotomy matters for an important reason. It converts a seemingly analytic, even topological, question into a finite, checkable computation. The question might be: does the function extend continuously? Does it blow up? Does it oscillate wildly? The answer just requires reading off the negative-index coefficients of a Laurent expansion. Once a singularity is classified, an entire toolkit becomes available. This ranges from Riemann's extension theorem to the residue calculus. Residue calculus computes contour integrals by bookkeeping local data at isolated points. The classification of singularities is the hinge on which the whole theory of meromorphic functions turns. The big picture is this: a holomorphic function on a punctured neighborhood of a point aa has a unique Laurent expansion there. The principal part of that expansion — the sum of terms with negative powers of (z−a)(z-a) — completely determines the type of singularity. If the principal part vanishes, the singularity is removable. The function secretly extends holomorphically. If the principal part is a finite nonzero sum, the function has a pole. It genuinely blows up, but in a controlled, algebraic way, the way 1/zn1/z^n blows up. If the principal part is an infinite sum, the singularity is essential. The function's local behavior is then as wild as possible. This fact is made vivid by the Casorati–Weierstrass theorem. It is made even more vivid by the Great Picard theorem. This article gives the precise definitions. It states and proves the classification theorem: Riemann's removable singularity theorem together with the pole and essential singularity criteria. It works through concrete examples with nontrivial coefficients. It shows exactly why each hypothesis cannot be dropped — isolatedness, boundedness, the punctured domain. It closes by pointing toward the Residue Theorem and the Argument Principle and Rouché's Theorem. These are the two great applications built on this classification.

Key takeaways
  • An isolated singularity of ff at aa means ff is holomorphic on some punctured disk 0<∣z−a∣<r0<|z-a|<r but not necessarily at aa itself.
  • Every such ff has a unique Laurent expansion f(z)=∑n=−∞∞cn(z−a)nf(z)=\sum_{n=-\infty}^\infty c_n (z-a)^n on the punctured disk.
  • The principal part ∑n<0cn(z−a)n\sum_{n<0} c_n(z-a)^n determines the type: zero part ⇒\Rightarrow removable, finitely many nonzero terms ⇒\Rightarrow pole of order =max⁡{−n:cn≠0}=\max\{-n : c_n\neq 0\}, infinitely many nonzero terms ⇒\Rightarrow essential.
  • Riemann's theorem: a singularity is removable iff ff is bounded near aa (equivalently iff lim⁡z→a(z−a)f(z)=0\lim_{z\to a} (z-a)f(z)=0).
  • Near a pole ∣f(z)∣→∞|f(z)|\to\infty; near an essential singularity, ff's image of every punctured neighborhood is dense in C\mathbb{C} (Casorati–Weierstrass), and in fact misses at most one value (Picard).
  • Key idea: the negative-power tail of a Laurent series is not just bookkeeping — it is a complete, computable invariant that sorts every isolated singularity into exactly one of three rigid categories.

Where this lives today

The classification of singularities is the load-bearing wall beneath the Residue Theorem: computing ∮γf dz\oint_\gamma f\,dz reduces to summing residues c−1c_{-1} at enclosed isolated singularities, and one must first know a singularity is a pole (not essential) to use the simple residue formulas (e.g. Res=lim⁡z→a(z−a)f(z)\text{Res} = \lim_{z\to a}(z-a)f(z) for a simple pole). It similarly underlies the Argument Principle and Rouché's Theorem, where the order of poles and zeros — read off exactly as in this classification — counts winding numbers of ff around 00. In physics, poles of Green's functions and propagators encode particle masses and resonances, while essential singularities appear in the study of the gamma function's reciprocal and in asymptotic/WKB analysis where Stokes phenomena occur. In engineering, the poles of a transfer function (Laplace-transform methods in control theory and signal processing) determine system stability, with pole order controlling the polynomial growth rate of impulse responses. In number theory and analytic combinatorics, the nature of a generating function's singularities (poles vs. essential singularities on the circle of convergence) governs the asymptotic growth rate of its coefficients, a cornerstone of the "singularity analysis" method for counting problems.

Ask the tutor: How does knowing a singularity is a pole (versus essential) let you compute its residue by a finite formula instead of a full Laurent expansion?

DefinitionIsolated singularity and its classification

Let a∈Ca\in\mathbb{C} and let ff be a holomorphic function on a punctured disk

D′(a,r)={z∈C:0<∣z−a∣<r}D'(a,r) = \{z\in\mathbb{C} : 0 < |z-a| < r\}

for some r>0r>0. We say ff has an isolated singularity at aa (we do not require ff to be defined at aa itself). By the Laurent expansion theorem, ff has a unique representation

f(z)=∑n=−∞∞cn(z−a)n,0<∣z−a∣<r,f(z) = \sum_{n=-\infty}^{\infty} c_n (z-a)^n, \qquad 0<|z-a|<r,

convergent absolutely and locally uniformly on D′(a,r)D'(a,r), where

cn=12πi∫∣z−a∣=ρf(z)(z−a)n+1 dz,0<ρ<r.c_n = \frac{1}{2\pi i}\int_{|z-a|=\rho} \frac{f(z)}{(z-a)^{n+1}}\,dz, \qquad 0<\rho<r.

The series ∑n=1∞c−n(z−a)−n\sum_{n=1}^\infty c_{-n}(z-a)^{-n} is called the principal part of ff at aa. The isolated singularity aa is classified as: 1. Removable if cn=0c_n = 0 for all n<0n<0 (the principal part vanishes); 2. A pole of order m≥1m\geq 1 if there is an integer m≥1m\geq 1 with c−m≠0c_{-m}\neq 0 and cn=0c_n=0 for all n<−mn<-m (the principal part is a nonzero finite sum, terminating exactly at index −m-m); 3. Essential if cn≠0c_n\neq 0 for infinitely many n<0n<0 (the principal part is an infinite series). These three cases are exhaustive and mutually exclusive, since the coefficients {cn}n<0\{c_n\}_{n<0} either are all zero, are eventually zero but not all zero, or are not eventually zero.

Intuition

The Laurent coefficients cnc_n for n<0n<0 measure how "singular" ff is, one power at a time. Think of c−1,c−2,c−3,…c_{-1}, c_{-2}, c_{-3},\dots as a signal: if the signal is entirely off, the function is secretly nice at aa and the puncture was cosmetic. If the signal switches off after finitely many terms — say only c−1c_{-1} and c−2c_{-2} are nonzero — the function behaves near aa exactly like a rational function c−2(z−a)−2+c−1(z−a)−1c_{-2}(z-a)^{-2}+c_{-1}(z-a)^{-1} plus something harmless, and it blows up to ∞\infty in a controlled polynomial way. If the signal never switches off, there is no finite algebraic model for the blow-up, and the function's behavior becomes essentially unpredictable: it does not tend to any limit (finite or infinite) as z→az\to a, and its values swarm densely around every complex number. A good mental picture is the map z↦1/zz\mapsto 1/z near z=0z=0: as ∣z∣→0|z|\to 0, ∣f(z)∣→∞|f(z)|\to\infty uniformly — this is the paradigm of a pole. Contrast this with z↦e1/zz\mapsto e^{1/z} near z=0z=0: along the positive real axis e1/z→∞e^{1/z}\to\infty, along the negative real axis e1/z→0e^{1/z}\to 0, and along the imaginary axis e1/ze^{1/z} oscillates on the unit circle. No single limit exists — the paradigm of an essential singularity.

Figure
Three principal-part signatures on the negative-index axis: zero tail (removable), finite tail (pole of order m=3m=3), infinite tail (essential)
Coefficients $c_{-n}$ for $n=1,2,3,4,5,\dots$ Removable: $c_{-1}=c_{-2}=\dots=0$ 00 00 0 Pole of order 3: $c_{-3}\neq 0$, then all zero $c_{-1}$ $c_{-2}$ $c_{-3}\ne0$ 0 0 Essential: infinitely many $c_{-n}\ne 0$ $\cdots$ never terminates
TheoremClassification of isolated singularities via Laurent data

Let ff be holomorphic on the punctured disk D′(a,r)={0<∣z−a∣<r}D'(a,r)=\{0<|z-a|<r\}, with Laurent expansion f(z)=∑n∈Zcn(z−a)nf(z)=\sum_{n\in\mathbb{Z}} c_n(z-a)^n. Then exactly one of the following holds, and each is equivalent to the stated analytic characterization.

(a) Removable. The following are equivalent: (i) cn=0c_n=0 for all n<0n<0; (ii) ff extends to a function holomorphic on the full disk D(a,r)D(a,r); (iii) ff is bounded on some punctured subdisk D′(a,ρ)D'(a,\rho), 0<ρ≤r0<\rho\le r; (iv) lim⁡z→a(z−a)f(z)=0\lim_{z\to a}(z-a)f(z)=0.

(b) Pole of order mm. The following are equivalent, for a fixed integer m≥1m\geq 1: (i) c−m≠0c_{-m}\neq0 and cn=0c_n=0 for n<−mn<-m; (ii) f(z)=(z−a)−mg(z)f(z) = (z-a)^{-m}g(z) on D′(a,r)D'(a,r) for some gg holomorphic on D(a,r)D(a,r) with g(a)≠0g(a)\neq 0; (iii) lim⁡z→a∣f(z)∣=∞\lim_{z\to a}|f(z)| = \infty, and mm is the unique positive integer such that lim⁡z→a(z−a)mf(z)\lim_{z\to a}(z-a)^m f(z) exists and is nonzero and finite.

(c) Essential. The following are equivalent: (i) cn≠0c_n\neq 0 for infinitely many n<0n<0; (ii) ff is neither removable nor a pole at aa; (iii) (Casorati–Weierstrass) for every ρ∈(0,r)\rho\in(0,r), the image f(D′(a,ρ))f(D'(a,\rho)) is dense in C\mathbb{C}.

Proof

We use the Laurent expansion (assumed known) as the starting analytic tool, together with basic estimates on Laurent coefficients and elementary limit arguments. The proof proceeds case by case; within each case, the equivalences are proved as a cycle of implications, which is logically airtight and avoids redundancy.

Preliminary estimate. For 0<ρ<r0<\rho<r, the Laurent coefficient formula gives, by the standard MLML-estimate on the circle ∣z−a∣=ρ|z-a|=\rho,

∣cn∣≤ρ−n M(ρ),M(ρ):=max⁡∣z−a∣=ρ∣f(z)∣.|c_n| \le \rho^{-n}\, M(\rho), \qquad M(\rho):=\max_{|z-a|=\rho}|f(z)|.

This will be used repeatedly to bound coefficients using boundedness of ff.

(a) Removable singularity. (i) ⇒\Rightarrow (ii): If cn=0c_n=0 for all n<0n<0, the Laurent series reduces to the power series ∑n≥0cn(z−a)n\sum_{n\geq0}c_n(z-a)^n, which converges on all of D(a,r)D(a,r) (Laurent series converge on their annulus of definition, and here the "inner radius" of convergence for the non-negative part extends to include aa since there is no singular part forcing an inner boundary). Define f~(z)=∑n≥0cn(z−a)n\tilde f(z)=\sum_{n\ge0}c_n(z-a)^n on D(a,r)D(a,r); this is holomorphic there (power series are holomorphic on their disk of convergence) and agrees with ff on D′(a,r)D'(a,r). Thus f~\tilde f is the desired holomorphic extension. (ii) ⇒\Rightarrow (iii): If ff extends to a holomorphic f~\tilde f on D(a,r)D(a,r), then f~\tilde f is continuous at aa, hence bounded on some closed subdisk D(a,ρ)‾⊂D(a,r)\overline{D(a,\rho)}\subset D(a,r) by continuity on a compact set. In particular f=f~f=\tilde f is bounded on D′(a,ρ)D'(a,\rho). (iii) ⇒\Rightarrow (iv): Suppose ∣f(z)∣≤K|f(z)|\le K on D′(a,ρ)D'(a,\rho). Then for z≠az\ne a with ∣z−a∣<ρ|z-a|<\rho,

∣(z−a)f(z)∣≤K∣z−a∣→0 as z→a.|(z-a)f(z)| \le K|z-a| \to 0 \text{ as } z\to a.

(iv) ⇒\Rightarrow (i): Suppose lim⁡z→a(z−a)f(z)=0\lim_{z\to a}(z-a)f(z)=0. In particular (z−a)f(z)(z-a)f(z) is bounded near aa, say ∣(z−a)f(z)∣≤K|(z-a)f(z)|\le K for 0<∣z−a∣<ρ′≤ρ0<|z-a|<\rho'\le\rho, so ∣f(z)∣≤K/∣z−a∣|f(z)|\le K/|z-a| there, giving M(ρ)≤K/ρM(\rho)\le K/\rho for all small ρ\rho. By the preliminary estimate, for each fixed n<0n<0,

∣cn∣≤ρ−nM(ρ)≤ρ−n⋅Kρ=Kρ−n−1.|c_n| \le \rho^{-n} M(\rho) \le \rho^{-n}\cdot\frac{K}{\rho} = K\rho^{-n-1}.

Since n<0n<0, we have −n−1≥0-n-1 \ge 0; write n=−kn=-k with k≥1k\ge1, so the bound is Kρk−1K\rho^{k-1}. If k≥2k\ge 2 this already →0\to 0 as ρ→0+\rho\to0^+, forcing cn=0c_n=0. For k=1k=1 (i.e. n=−1n=-1) we need the sharper hypothesis (iv) itself: given ε>0\varepsilon>0, choose ρ\rho small enough that ∣(z−a)f(z)∣<ε|(z-a)f(z)|<\varepsilon for ∣z−a∣=ρ|z-a|=\rho; then M(ρ)⋅ρ<εM(\rho)\cdot \rho <\varepsilon on that circle is not quite the right form, so redo the estimate directly: from ∣(z−a)f(z)∣≤ε(ρ)|(z-a)f(z)|\le \varepsilon(\rho) on ∣z−a∣=ρ|z-a|=\rho with ε(ρ)→0\varepsilon(\rho)\to0, we get M(ρ)≤ε(ρ)/ρM(\rho)\le \varepsilon(\rho)/\rho, hence

∣c−1∣≤ρ1⋅M(ρ)≤ρ⋅ε(ρ)ρ=ε(ρ)→0,|c_{-1}| \le \rho^{1}\cdot M(\rho) \le \rho\cdot\frac{\varepsilon(\rho)}{\rho} = \varepsilon(\rho) \to 0,

so c−1=0c_{-1}=0. Combined with the k≥2k\ge2 case, all cn=0c_n=0 for n<0n<0. This closes the cycle (i)⇒\Rightarrow(ii)⇒\Rightarrow(iii)⇒\Rightarrow(iv)⇒\Rightarrow(i), proving all four are equivalent. This equivalence is exactly Riemann's removable singularity theorem.

(b) Pole of order mm. (i) ⇒\Rightarrow (ii): If c−m≠0c_{-m}\ne0 and cn=0c_n=0 for n<−mn<-m, write

f(z)=∑n=−m∞cn(z−a)n=(z−a)−m∑n=−m∞cn(z−a)n+m=(z−a)−mg(z),f(z) = \sum_{n=-m}^{\infty} c_n (z-a)^n = (z-a)^{-m}\sum_{n=-m}^\infty c_n(z-a)^{n+m} = (z-a)^{-m} g(z),

where g(z):=∑k=0∞ck−m(z−a)kg(z):=\sum_{k=0}^\infty c_{k-m}(z-a)^k (reindexing k=n+m≥0k=n+m\ge0). This is a power series with the same radius of convergence as the original Laurent series' outer part, hence converges and is holomorphic on all of D(a,r)D(a,r), and g(a)=c−m≠0g(a)=c_{-m}\ne0. (ii) ⇒\Rightarrow (iii): Since gg is holomorphic at aa with g(a)≠0g(a)\ne0, continuity gives g(z)→g(a)≠0g(z)\to g(a)\ne0 as z→az\to a, and gg is bounded away from 00 near aa (by continuity, ∣g(z)∣>12∣g(a)∣>0|g(z)|>\tfrac12|g(a)|>0 on some punctured disk around aa). Then

∣f(z)∣=∣g(z)∣∣z−a∣m→∞as z→a,|f(z)| = \frac{|g(z)|}{|z-a|^m} \to \infty \quad \text{as } z\to a,

since the numerator is bounded below by a positive constant and the denominator →0\to0. Moreover,

(z−a)mf(z)=g(z)→g(a)≠0, finite,(z-a)^m f(z) = g(z) \to g(a) \ne 0, \text{ finite},

so the limit lim⁡z→a(z−a)mf(z)\lim_{z\to a}(z-a)^m f(z) exists, is finite and nonzero. For uniqueness of such mm: if j<mj<m then (z−a)jf(z)=(z−a)j−mg(z)→∞(z-a)^j f(z) = (z-a)^{j-m}g(z)\to\infty (since j−m<0j-m<0 and g(a)≠0g(a)\ne0), so the limit does not exist finitely; if j>mj>m then (z−a)jf(z)=(z−a)j−mg(z)→0(z-a)^jf(z) = (z-a)^{j-m}g(z)\to 0, giving the wrong (zero) limit. So mm is indeed the unique exponent giving a finite nonzero limit. (iii) ⇒\Rightarrow (i): Suppose ∣f(z)∣→∞|f(z)|\to\infty as z→az\to a and mm is the unique positive integer with L:=lim⁡z→a(z−a)mf(z)L:=\lim_{z\to a}(z-a)^mf(z) finite and nonzero. Since (z−a)mf(z)→L≠0(z-a)^mf(z)\to L\ne0 finite, this function is bounded near aa: there exist ρ0>0,K>0\rho_0>0,K>0 with ∣(z−a)mf(z)∣≤K|(z-a)^m f(z)|\le K for 0<∣z−a∣<ρ00<|z-a|<\rho_0. Hence ∣f(z)∣≤K∣z−a∣−m|f(z)|\le K|z-a|^{-m}, so M(ρ)≤Kρ−mM(\rho)\le K\rho^{-m} for small ρ\rho. By the preliminary estimate, for n<−mn<-m, writing n=−m−jn=-m-j with j≥1j\ge1,

∣cn∣≤ρ−nM(ρ)≤ρm+j⋅Kρ−m=Kρj→0 as ρ→0+,|c_n| \le \rho^{-n}M(\rho) \le \rho^{m+j}\cdot K\rho^{-m} = K\rho^{j} \to 0 \text{ as }\rho\to0^+,

forcing cn=0c_n=0 for all n<−mn<-m. It remains to show c−m≠0c_{-m}\ne0. Since (z−a)mf(z)→L≠0(z-a)^mf(z)\to L\ne 0 and (from the vanishing of cnc_n, n<−mn<-m, just shown) (z−a)mf(z)=∑k≥0ck−m(z−a)k(z-a)^m f(z) = \sum_{k\ge0}c_{k-m}(z-a)^k is itself a convergent power series near aa whose value at z=az=a (by continuity of power series) equals its constant term c−mc_{-m}; matching with the limit LL gives c−m=L≠0c_{-m}=L\ne0. This closes the cycle (i)⇒\Rightarrow(ii)⇒\Rightarrow(iii)⇒\Rightarrow(i).

(c) Essential singularity. (i) ⇔\Leftrightarrow (ii): By definition, the singularity is removable iff all cn=0c_n=0 (n<0n<0), a pole iff finitely many (but at least one) cn≠0c_n\ne0 for n<0n<0 terminating at some −m-m. Since the three cases (all zero / eventually zero, not all / not eventually zero) exhaust and are mutually exclusive for the coefficient sequence (cn)n<0(c_n)_{n<0}, "essential" (infinitely many nonzero cnc_n, n<0n<0) is logically identical to "neither removable nor a pole." This is immediate from the definition and requires no further analytic input. (ii) ⇒\Rightarrow (iii) (Casorati–Weierstrass): Suppose, for contradiction, that for some ρ∈(0,r)\rho\in(0,r) the image f(D′(a,ρ))f(D'(a,\rho)) is not dense in C\mathbb{C}. Then there exists w0∈Cw_0\in\mathbb{C} and δ>0\delta>0 such that

∣f(z)−w0∣≥δfor all z∈D′(a,ρ).|f(z) - w_0| \ge \delta \quad \text{for all } z\in D'(a,\rho).

Define h(z):=1f(z)−w0h(z) := \dfrac{1}{f(z)-w_0} on D′(a,ρ)D'(a,\rho). Then hh is holomorphic there (as f−w0≠0f-w_0\ne0 and holomorphic), and ∣h(z)∣≤1/δ|h(z)| \le 1/\delta for all z∈D′(a,ρ)z\in D'(a,\rho), so hh is bounded on a punctured disk around aa. By part (a) [the equivalence (iii)⇒\Rightarrow(i) already proved, i.e. Riemann's theorem], hh has a removable singularity at aa; let h~\tilde h be its holomorphic extension to D(a,ρ)D(a,\rho). Case 1: h~(a)≠0\tilde h(a) \ne 0. Then near aa, f(z)=w0+1h(z)=w0+1h~(z)f(z) = w_0 + \dfrac{1}{h(z)} = w_0 + \dfrac{1}{\tilde h(z)}, and since h~\tilde h is holomorphic and nonvanishing near aa (by continuity, as h~(a)≠0\tilde h(a)\ne0), the right side is holomorphic near aa. So ff extends holomorphically to aa, i.e. ff has a removable singularity — contradicting (ii). Case 2: h~(a)=0\tilde h(a) = 0. Since h~\tilde h is holomorphic and not identically zero near aa (it is nonzero on the punctured disk, as h=1/(f−w0)h=1/(f-w_0) never vanishes there), the zero at aa has some finite order k≥1k\ge1: h~(z)=(z−a)ku(z)\tilde h(z) = (z-a)^k u(z) with uu holomorphic and u(a)≠0u(a)\ne0 near aa (this is the standard local factorization of zeros of a nonconstant holomorphic function, itself following from the Taylor expansion of h~\tilde h at aa). Then near aa,

f(z)−w0=1h~(z)=1(z−a)ku(z),f(z) - w_0 = \frac{1}{\tilde h(z)} = \frac{1}{(z-a)^k u(z)},

so (z−a)k(f(z)−w0)=1/u(z)→1/u(a)≠0(z-a)^k(f(z)-w_0) = 1/u(z) \to 1/u(a) \ne 0, finite. This is exactly criterion (iii) of part (b) (applied to f−w0f - w_0, which has the same singularity type as ff since w0w_0 is a constant shift and does not change the principal part), showing ff has a pole of order kk at aa — again contradicting (ii). Both cases contradict the hypothesis that ff is neither removable nor a pole at aa. Hence the assumption that f(D′(a,ρ))f(D'(a,\rho)) fails to be dense must be false, for every ρ∈(0,r)\rho \in (0,r). This proves (iii). (iii) ⇒\Rightarrow (ii): We prove the contrapositive. Suppose ff is removable or a pole at aa. If removable, ff extends continuously (indeed holomorphically) to aa, so f(z)→f(a)f(z)\to f(a) as z→az\to a; then for ρ\rho small, f(D′(a,ρ))f(D'(a,\rho)) lies in a small disk around f(a)f(a) and is certainly not dense in C\mathbb{C}. If ff has a pole at aa, by part (b) ∣f(z)∣→∞|f(z)|\to\infty as z→az\to a, so for ρ\rho small, f(D′(a,ρ))f(D'(a,\rho)) lies outside a large disk {∣w∣≤K}\{|w|\le K\} and again is not dense in C\mathbb{C} (it misses the entire disk {∣w∣<K}\{|w|<K\}). Either way (iii) fails for small ρ\rho. This proves the contrapositive, hence (iii)⇒\Rightarrow(ii). This closes the logical cycle for part (c) and completes the proof of the theorem. ■\blacksquare

∎

Worked examples

Example 1 — a rational function with an asymmetric pole structure. Let

f(z)=z2−3z+5(z−2)3.f(z) = \frac{z^2 - 3z + 5}{(z-2)^3}.

We classify the singularity at a=2a=2 directly from the definitions, without appealing to a general "order of pole of a rational function" shortcut, to illustrate the Laurent-coefficient method concretely. Substitute w=z−2w=z-2, so z=w+2z=w+2:

z2−3z+5=(w+2)2−3(w+2)+5=w2+4w+4−3w−6+5=w2+w+3.z^2-3z+5 = (w+2)^2 - 3(w+2) + 5 = w^2+4w+4 -3w-6+5 = w^2 + w + 3.

Hence

f(z)=w2+w+3w3=3w−3+w−2+w−1,w=z−2≠0.f(z) = \frac{w^2+w+3}{w^3} = 3w^{-3} + w^{-2} + w^{-1}, \qquad w = z-2 \ne 0.

This is already the Laurent expansion at a=2a=2 (it has finitely many terms, all with nonzero coefficients down to w−3w^{-3}, and no non-negative powers at all): c−3=3c_{-3}=3, c−2=1c_{-2}=1, c−1=1c_{-1}=1, and cn=0c_n=0 otherwise. The most negative nonzero index is −3-3, so by the classification theorem part (b), ff has a pole of order 3 at z=2z=2. Consistency check via criterion (iii): (z−2)3f(z)=z2−3z+5→4−6+5=3≠0(z-2)^3f(z) = z^2-3z+5 \to 4-6+5=3\ne0 as z→2z\to2, confirming order 33 with the correct nonzero limit L=3=c−3L=3=c_{-3}, matching the coefficient found directly.

Example 2 — telling apart three functions built from the same building block. Consider the three functions, all singular at a=0a=0:

f1(z)=sin⁡zz,f2(z)=cos⁡z−1z2,f3(z)=e1/z2sin⁡ ⁣(1z).f_1(z) = \frac{\sin z}{z}, \qquad f_2(z) = \frac{\cos z - 1}{z^2}, \qquad f_3(z) = e^{1/z^2}\sin\!\left(\frac{1}{z}\right).

For f1f_1: using the known Taylor series sin⁡z=z−z36+z5120−⋯\sin z = z - \dfrac{z^3}{6} + \dfrac{z^5}{120} - \cdots (valid for all z∈Cz\in\mathbb{C}), divide by zz:

f1(z)=1−z26+z4120−⋯ ,z≠0.f_1(z) = 1 - \frac{z^2}{6} + \frac{z^4}{120} - \cdots, \qquad z\ne 0.

This is a Laurent series with no negative powers at all (cn=0c_n=0 for n<0n<0). By part (a)(i)⇒\Rightarrow(ii), z=0z=0 is a removable singularity, and the extension is f~1(0)=1\tilde f_1(0)=1 (the constant term), matching the classical fact lim⁡z→0sin⁡z/z=1\lim_{z\to0}\sin z/z = 1. For f2f_2: using cos⁡z=1−z22+z424−z6720+⋯\cos z = 1 - \dfrac{z^2}{2} + \dfrac{z^4}{24} - \dfrac{z^6}{720}+\cdots, we get cos⁡z−1=−z22+z424−⋯\cos z - 1 = -\dfrac{z^2}{2}+\dfrac{z^4}{24}-\cdots, so dividing by z2z^2,

f2(z)=−12+z224−z4720+⋯ ,z≠0.f_2(z) = -\frac12 + \frac{z^2}{24} - \frac{z^4}{720} + \cdots, \qquad z\ne0.

Again cn=0c_n=0 for n<0n<0: removable, with extension value f~2(0)=−12\tilde f_2(0) = -\tfrac12. For f3f_3: this is the interesting case. Using eu=∑k≥0uk/k!e^{u} = \sum_{k\ge0} u^k/k! with u=1/z2u=1/z^2, and sin⁡v=∑j≥0(−1)jv2j+1/(2j+1)!\sin v = \sum_{j\ge0}(-1)^j v^{2j+1}/(2j+1)! with v=1/zv=1/z, both series converge for all z≠0z\ne0 (entire functions composed with 1/z1/z), and their product, rearranged as a Laurent series in powers of 1/z1/z (justified because both series are absolutely convergent for each fixed z≠0z\ne 0, so their Cauchy product may be freely rearranged), has the form

f3(z)=∑ncnz−n,n≥1,f_3(z) = \sum_{n} c_n z^{-n}, \qquad n \ge 1,

with no non-negative powers of zz appearing at all (since e1/z2e^{1/z^2} contributes only even non-positive powers z−2kz^{-2k}, k≥0k\ge0, and sin⁡(1/z)\sin(1/z) contributes only odd negative powers z−(2j+1)z^{-(2j+1)}, j≥0j\ge0; every product term z−2k⋅z−(2j+1)=z−(2k+2j+1)z^{-2k}\cdot z^{-(2j+1)} = z^{-(2k+2j+1)} has a strictly negative, odd exponent, and different pairs (k,j)(k,j) can coincide in exponent but never cancel completely to zero because the leading term, k=j=0k=j=0, gives z−1z^{-1} with coefficient 1⋅1=1≠01\cdot1=1\ne0, and one checks inductively that infinitely many such exponents −(2k+2j+1)-(2k+2j+1) occur with nonzero coefficient sums as k,jk,j range over all nonnegative integers). Concretely the first few terms are

f3(z)=1z+1z3+12z5+⋯ ,f_3(z) = \frac{1}{z} + \frac{1}{z^3} + \frac{1}{2z^5} + \cdots,

(from k=0,j=0k=0,j=0: coefficient of z−1z^{-1} is 11; from k=1,j=0k=1,j=0 and k=0,j=1k=0,j=1 combined at exponent −3-3: 1⋅11!⋅−13!∣j=1+12!⋅1∣k=11\cdot\frac{1}{1!}\cdot\frac{-1}{3!}\big|_{j=1} + \frac{1}{2!}\cdot 1\big|_{k=1}, working out to 11!⋅1−11!⋅16+12⋅1=12\frac{1}{1!}\cdot 1 - \frac1{1!}\cdot\frac16 + \frac1{2}\cdot 1 = \frac12 after collecting; the exact combinatorics is routine Cauchy-product bookkeeping and the key structural point — infinitely many nonzero negative-index coefficients — is what matters). Since infinitely many negative-power coefficients are nonzero (there is no finite stopping point, as the exponents −(2k+2j+1)-(2k+2j+1) take every odd negative integer value infinitely often as k,j→∞k,j\to\infty, and no infinite cancellation can force all but finitely many of these sums to vanish for an analytic function unless f3f_3 were a pole/removable type, which we can also confirm by the Casorati–Weierstrass test below), z=0z=0 is an essential singularity of f3f_3. As an independent check consistent with part (c)(iii): along the real axis, sin⁡(1/x)\sin(1/x) oscillates in [−1,1][-1,1] while e1/x2→+∞e^{1/x^2}\to+\infty as x→0x\to0, so f3f_3 is unbounded near 00 along most directions but can be made close to 00 along the sequence xn=1/(nπ)x_n = 1/(n\pi) where sin⁡(1/xn)=sin⁡(nπ)=0\sin(1/x_n)=\sin(n\pi)=0 exactly — so f3(xn)=0f_3(x_n)=0 for every nn, showing f3f_3 does not tend to ∞\infty, ruling out a pole, and combined with unboundedness, ruling out removability. This is exactly the erratic behavior the classification theorem predicts for essential singularities.

Counterexampledropping "isolated" breaks the trichotomy

The classification theorem is stated for an isolated singularity: ff holomorphic on a full punctured disk 0<∣z−a∣<r0<|z-a|<r. If singularities are allowed to accumulate at aa, none of the three clean categories need apply, and there may not even be a Laurent expansion valid on a punctured disk around aa at all. Consider

f(z)=1sin⁡(1/z),z≠0, z≠1nπ (n∈Z∖{0}).f(z) = \frac{1}{\sin(1/z)}, \qquad z\ne 0, \ z \ne \frac{1}{n\pi} \ (n\in\mathbb{Z}\setminus\{0\}).

The function sin⁡(1/z)\sin(1/z) vanishes exactly at the points zn=1/(nπ)z_n = 1/(n\pi) for every nonzero integer nn, and these points accumulate at z=0z=0 (since zn→0z_n\to 0 as ∣n∣→∞|n|\to\infty). At each such znz_n, ff has a pole (since sin⁡(1/z)\sin(1/z) has a simple zero there, as one checks via the chain rule: ddzsin⁡(1/z)∣z=zn=−1zn2cos⁡(nπ)=−(nπ)2(−1)n1≠0\frac{d}{dz}\sin(1/z)\big|_{z=z_n} = -\frac{1}{z_n^2}\cos(n\pi) = -\frac{(n\pi)^2(-1)^n}{1}\ne0, a nonzero derivative, so the zero of sin⁡(1/z)\sin(1/z) at znz_n is simple, giving ff a simple pole there by part (b)(ii) applied locally). But no punctured disk 0<∣z∣<r0<|z|<r around 00 avoids all the points znz_n, because zn→0z_n\to0: for any r>0r>0, choosing nn with 1/(nπ)<r1/(n\pi) < r places znz_n inside the punctured disk. Hence ff is not holomorphic on any punctured disk D′(0,r)D'(0,r) — the hypothesis of an isolated singularity fails outright, since the singularities of ff near 00 are not isolated from each other. There is consequently no Laurent series for ff valid on any annulus 0<∣z∣<r0<|z|<r, and the trichotomy (removable / pole / essential) simply does not apply to the point z=0z=0: it is a non-isolated singularity, sometimes called a limit point of poles, and it falls entirely outside the theorem's scope. This shows why "isolated" cannot be weakened to merely "ff is holomorphic on a deleted neighborhood minus some singular set accumulating at aa" — the very existence of the Laurent expansion, on which the whole classification rests, requires an honest punctured disk free of any other singularities.

Scope, edge cases & common traps

SituationWhat to checkCommon error
ff has a removable singularityCompute lim⁡z→a(z−a)f(z)\lim_{z\to a}(z-a)f(z); must be 00Confusing "bounded" with "has a limit" — boundedness alone (part (a)(iii)) already suffices, no need to separately verify existence of lim⁡f(z)\lim f(z)
ff has a poleFind the unique mm with (z−a)mf(z)→L≠0(z-a)^mf(z)\to L\ne0 finiteGuessing mm from a formula for ff without verifying c−m≠0c_{-m}\ne0, e.g. mis-cancelling a numerator zero against a denominator zero
ff has an essential singularityConfirm infinitely many negative-index Laurent coefficients are nonzeroConcluding "essential" merely because a limit "does not exist," without checking it isn't oscillating between finitely many poles (see non-isolated counterexample)
Composite/product functionsLaurent series must be justified by absolute convergence before rearranging as a Cauchy productFormally multiplying two Laurent series in 1/z1/z without justifying convergence/rearrangement
Non-isolated singular pointsVerify a genuine punctured disk exists with no other singularities of ffApplying the trichotomy to accumulation points of poles, essential singularities, or branch points

The mathematician behind it

Origins · 1 minBernhard Riemann and how this was discovered
BR
Bernhard Riemann
1826–1866 · Göttingen, mid-19th-century complex analysis

Riemann's 1851 doctoral dissertation, supervised by Gauss, laid the geometric foundations of complex analysis: conformal mappings, what became Riemann surfaces, and the removable singularity criterion that today bears his name. His approach fused analysis with geometric and topological intuition — thinking of holomorphic functions as maps between surfaces rather than mere formal power series — a perspective that reshaped the entire field and underlies the modern treatment of singularities as local invariants detectable by growth conditions like boundedness.

History
- 1821 — Cauchy develops the foundations of complex integration and the residue idea in early form.
- 1851 — Riemann's dissertation introduces Riemann surfaces and conformal mapping, with the removable singularity theorem as a key local tool.
- 1876 — Weierstrass and Casorati (independently, building on earlier remarks by Weierstrass) establish the density theorem for essential singularities.
- 1879 — Picard proves the Great Picard theorem, sharpening Casorati–Weierstrass to show at most one exceptional value is omitted near an essential singularity.
- 1880s — Mittag-Leffler and others systematize the pole/principal-part calculus into the modern theory of meromorphic functions.
How do I compute the order of a pole quickly without a full Laurent expansion?

Write f(z)=g(z)/(z−a)mf(z) = g(z)/(z-a)^m where gg is holomorphic near aa with g(a)≠0g(a)\ne 0; then aa is a pole of order exactly mm. Equivalently, compute lim⁡z→a(z−a)kf(z)\lim_{z\to a}(z-a)^k f(z) for successive integers k=1,2,…k=1,2,\dots: the order is the unique kk where this limit is finite and nonzero (part (b)(iii) of the theorem). For rational functions this reduces to factoring the denominator and cancelling common factors with the numerator before reading off multiplicities — cancelling first is essential, as illustrated in Example 1's companion trap in the scope table.

Is every essential singularity as wild as Casorati–Weierstrass suggests, or can it be worse?

Casorati–Weierstrass only guarantees the image is dense; the much stronger Great Picard theorem (1879) shows that in every punctured neighborhood of an essential singularity, ff attains every complex value infinitely often, with at most one possible exception. For e1/ze^{1/z} at 00, the omitted value is exactly 00, since e1/ze^{1/z} never equals 00 for any finite z≠0z\ne 0; every other complex number is hit infinitely often in any punctured disk around the origin. Picard's theorem is considerably deeper and uses tools (normal families, the modular function) well beyond Laurent series alone.

Can a function have different singularity types at different points?

Yes, and this is the typical situation for a general meromorphic or more complicated function. For example f(z)=1z+e1/(z−1)f(z) = \dfrac{1}{z} + e^{1/(z-1)} has a simple pole at z=0z=0 and an essential singularity at z=1z=1; each isolated singularity is analyzed completely independently using its own local Laurent expansion around that particular point, since the classification is a purely local invariant depending only on behavior in an arbitrarily small punctured disk around each singularity separately.

What happens at z=∞z=\infty — does the classification extend there?

Yes, by the standard device of setting w=1/zw=1/z and studying g(w)=f(1/w)g(w)=f(1/w) near w=0w=0; the singularity type of ff at ∞\infty is defined to be whatever type gg has at w=0w=0. For instance, any nonconstant polynomial p(z)p(z) has a pole at ∞\infty (of order equal to its degree), since p(1/w)p(1/w) has a pole at w=0w=0 of that order; eze^z has an essential singularity at ∞\infty, since e1/we^{1/w} is essential at w=0w=0. This convention is what makes the Riemann sphere C∪{∞}\mathbb{C}\cup\{\infty\} a natural domain for meromorphic function theory.

Why does boundedness alone force removability — isn't that surprising, given no continuity is assumed at aa?

It is one of the most striking rigidity phenomena in complex analysis: holomorphy on the punctured disk is an extremely strong constraint, far stronger than smoothness in real analysis (where, e.g., sin⁡(1/x)\sin(1/x) is bounded near 00 but has no continuous extension). The proof works because the Laurent coefficients cnc_n for n<0n<0 are computed by contour integrals over shrinking circles, and a uniform bound on ff forces these integrals — which pick up a factor ρ−n\rho^{-n} with −n>0-n>0 shrinking to 00 — to vanish in the limit ρ→0\rho\to 0. This ties analytic boundedness directly to the vanishing of specific algebraic coefficients, a link with no real-variable analogue.

How does this classification set up the Residue Theorem?

The residue theorem states that ∮γf dz=2πi∑Res(f,ak)\oint_\gamma f\,dz = 2\pi i \sum \text{Res}(f,a_k) summed over isolated singularities enclosed by γ\gamma, where Res(f,a):=c−1\text{Res}(f,a) := c_{-1}, the coefficient extracted directly from the classification data above. Removable singularities contribute residue 00 automatically (since c−1=0c_{-1}=0 there); poles have computable residues via explicit formulas involving derivatives of g(z)=(z−a)mf(z)g(z)=(z-a)^mf(z); essential singularities require the full Laurent expansion since no finite algebraic formula for c−1c_{-1} exists in general. Thus this classification is the prerequisite that tells you which formula for the residue is even applicable.

Does "isolated singularity" require ff to be undefined at aa, or just possibly badly behaved there?

The standard convention is that ff is holomorphic on the punctured disk D′(a,r)D'(a,r) and ff is simply not assumed to be defined (or, if defined, not assumed holomorphic) at aa itself. If ff actually happens to already be defined and holomorphic at aa, it trivially has a "removable" singularity there in a degenerate sense (nothing to remove), and the theorem's conclusions still hold, just vacuously. The interesting content is entirely for functions where aa is excluded from the natural domain of definition, such as sin⁡(z)/z\sin(z)/z at z=0z=0 before one notices the removable extension.

Before this

Where this leads