Analytic Continuation

A power series $\sum a_n (z-z_0)^n$ is a local object: it defines an analytic function only inside its disk of convergence, whose radius $R$ is the distance to the nearest singularity and no larger. Yet the function it names almost always lives on a far bigger set — the geometric series knows only $|z|<1$ , but the function it sums, $1/(1-z)$ , is holomorphic on all of $\mathbb{C}$ except a single point. Analytic continuation is both the procedure that recovers this larger function from a local seed and the theory that says when the recovery is unique. The engine is a phenomenon with no analogue in real calculus: rigidity. A holomorphic function is not free to be assigned values as it pleases. Once it is known on any set with an interior limit point — a tiny arc, a convergent sequence of points, an open disk however small — it is fixed on every connected open set reachable from that data. This is the identity theorem, and it is what makes "the" continuation unique whenever it exists. Constructively, one re-expands the Taylor series about a point near the rim of the current disk; the fresh disk may stick out past the old one, and the two representations agree on the overlap. Chaining such disks along a path realizes continuation along a curve, and when topologically distinct paths around a singular point disagree, we meet branch points, multivaluedness, and the monodromy theorem. In physics the technique is everywhere: dimensional regularization continues loop integrals in the spacetime dimension, and zeta regularization tames divergent vacuum sums.

Key takeaways

Analytic continuation extends a holomorphic function from its original disk of convergence to a larger domain by patching together overlapping local power-series representations.
The identity theorem makes the extension rigid and unique: two holomorphic functions that agree on any set with an interior limit point agree everywhere on a connected domain.
The basic move is re-expansion — a Taylor series about a point near the edge of one disk converges in a new disk that can poke beyond the old boundary, provided a singularity does not block that direction.
Continuation along homotopically inequivalent paths around a singularity can give different answers; this multivaluedness ( $\log z$ , $z^{1/2}$ ) is organized by the monodromy theorem and Riemann surfaces.
Physically it defines $\Gamma(z)$ , $\zeta(s)$ , and scattering amplitudes outside the region where their integral or series representations converge — and underlies dimensional and zeta regularization.
Some series (lacunary series like $\sum z^{2^n}$ ) admit no continuation at all: their circle of convergence is a natural boundary.
Key idea: holomorphy is rigid, so overlapping representations can only patch together in one consistent way — analytic continuation is the unique maximal extension along each homotopy class of paths, blocked only by singularities and natural boundaries.

The core insight

Holomorphy is rigid: by the identity theorem an analytic function on a connected domain is uniquely fixed by its values on any set with an interior limit point. Hence overlapping power-series disks (or any matching representations) can only patch together in one consistent way, defining a unique maximal extension along each homotopy class of paths.

Where this lives today

Analytic continuation is not a museum piece; it is the daily working tool of theoretical physics. In quantum field theory, dimensional regularization computes a loop integral in $d$ spacetime dimensions, obtains a closed form involving $\Gamma(n-d/2)$ , and then continues in $d$ to $d=4-\epsilon$ — isolating ultraviolet divergences as simple poles in $\epsilon$ that renormalization then absorbs. The entire scheme rests on continuing the Gamma function, and it is preferred because it preserves gauge and Lorentz symmetry where a hard cutoff would not. Zeta regularization uses $\zeta(-1)=-\tfrac1{12}$ and $\zeta'(0)$ to assign finite values to divergent determinants and vacuum sums, from the Casimir force to the critical dimension of string theory. In scattering theory, causality is analyticity: amplitudes continued into complex energy obey dispersion relations, and continuation in complex angular momentum (Regge theory) organizes high-energy hadronic cross sections. The frontier open problem is numerical: lattice QCD computes correlators in Euclidean (imaginary) time, and extracting real-time spectral functions — transport coefficients, particle widths — requires continuing noisy discrete data back to Minkowski signature, an ill-posed inversion tackled with the maximum-entropy method and Bayesian reconstruction. Taming that continuation is a live programme connecting the quark-gluon plasma to the mathematics Weierstrass began.

The physical picture

Picture a power series as a flashlight that lights only a circular disk of the complex plane; the analytic function is the whole dark room. Walk the flashlight to the rim of its beam and re-aim it — that is, re-expand the Taylor series about a point near the edge of the current disk — and it lights a new patch that overlaps the old one. Where two patches overlap they must agree, because the identity theorem forbids any other outcome. Patch by patch you map the entire room, stopped only by the walls: the singularities, which you cannot walk through. The first figure carries out exactly one such step on the geometric series and shows, with real numbers, how the fresh disk escapes the original unit circle.

Figure

Re-expanding the geometric series about z₁ = −1/2 produces a new disk of radius R₁ = 3/2 that reaches out to z = −2, past the original unit circle, while the pole at z = 1 stops it there — a single step of analytic continuation made visible

Rigidity is the genuine surprise. A merely smooth real function can be altered on a small interval and stitched back invisibly; a holomorphic function cannot be touched anywhere without changing it everywhere. Knowing $f$ on a tiny arc, or even on a sequence of points converging to an interior point, pins it down throughout the connected region. The information is non-local in the strongest possible sense — holomorphy threads every value to every other value. Multivaluedness, meanwhile, is geometry rather than pathology. If you carry the flashlight in a loop around a branch point, you can return to the starting position holding a different value: $\log z$ comes back larger by $2\pi i$ . The function is perfectly single-valued once you accept that it lives not on the plane but on a spiral staircase — a Riemann surface — of which the plane is only the shadow. The second figure follows the logarithm around the origin and stacks the sheets it climbs.

Figure

Continuing log z once counterclockwise around the branch point at z = 0 adds ∮ dz/z = 2πi, so log 1 climbs from 0 to 2πi — the logarithm is single-valued on an infinite spiral of sheets stacked in steps of 2πi

The core equations

Everything begins with rigidity, stated as the identity theorem. Two functions holomorphic on a connected domain that agree on a surprisingly thin set are forced to agree everywhere — this is the uniqueness engine of the whole subject.

f\equiv g \text{ on } \Omega \iff f=g \text{ on } S\subset\Omega \text{ with a limit point in } \Omega \qquad (1)

where: $\Omega$ — a connected open subset (domain) of $\mathbb{C}$ ; $f,g$ — functions holomorphic on $\Omega$ ; $S$ — any set on which $f$ and $g$ are known to agree, having a limit point interior to $\Omega$ . Reading: agreement on a mere sequence of points converging inside $\Omega$ already forces global agreement, so a continuation, once found, is the only one. The reach of a single series is fixed by the Cauchy–Hadamard formula: the radius of convergence equals the distance from the centre to the nearest singularity, and not one unit more. This is why one series can never see the whole function whenever singularities are present.

f(z)=\sum_{n=0}^{\infty} a_n (z-z_0)^n,\qquad R=\frac{1}{\limsup_{n\to\infty}|a_n|^{1/n}} \qquad (2)

where: $a_n=f^{(n)}(z_0)/n!$ — the Taylor coefficients; $z_0$ — centre of expansion; $R$ — radius of convergence, equal to the distance from $z_0$ to the nearest singularity of $f$ . Reading: the boundary of the disk always carries at least one singularity, and that singularity is what a continuation must route around. To escape the disk we re-expand about a nearby interior point $z_1$ . The new coefficients are computed from the old ones, and the new disk has its own radius set by the nearest singularity to $z_1$ — which may lie farther away in some directions.

g(z)=\sum_{n=0}^{\infty} b_n (z-z_1)^n,\qquad b_n=\frac{1}{n!}\sum_{k=n}^{\infty}\frac{k!}{(k-n)!}\,a_k\,(z_1-z_0)^{k-n} \qquad (3)

where: $z_1$ — new centre with $|z_1-z_0|<R$ ; $b_n$ — coefficients of the re-expanded series; $g$ — the continuation, which agrees with $f$ on the overlap $D_0\cap D_1$ . Reading: $b_n=f^{(n)}(z_1)/n!$ , so re-expansion is just Taylor's theorem applied at a shifted centre — no new information, only a new vantage point. Chaining such disks along a curve realizes continuation along a path, and the monodromy theorem controls when the answer depends only on the path's endpoints. In a simply connected domain free of singular obstructions, continuation is path-independent; around a branch point it is not.

\mathcal{C}:\ D_0\!\to\!D_1\!\to\!\cdots\!\to\!D_N,\qquad \text{homotopic paths in } \Omega \Rightarrow \text{same continuation} \qquad (4)

where: $\mathcal{C}$ — a curve covered by a finite chain of overlapping disks $D_j$ with $D_j\cap D_{j+1}\neq\varnothing$ ; $\Omega$ — a domain across which the homotopy avoids singularities. Reading: two paths that can be deformed into one another without crossing a singularity yield the same endpoint value; two that cannot may disagree. The archetype in physics is the Gamma function. Its integral converges only for $\Re z>0$ , but the functional equation continues it to a meromorphic function on all of $\mathbb{C}$ with simple poles at the non-positive integers.

\Gamma(z)=\int_0^{\infty} t^{z-1}e^{-t}\,\mathrm{d}t\ \ (\Re z>0),\qquad \Gamma(z)=\frac{\Gamma(z+n)}{z(z+1)\cdots(z+n-1)} \qquad (5)

where: $z$ — complex argument; $n$ — any integer with $\Re z+n>0$ ; the recurrence continues $\Gamma$ to $\mathbb{C}\setminus\{0,-1,-2,\dots\}$ , with a simple pole of residue $(-1)^n/n!$ at $z=-n$ . Reading: the recurrence trades a value in an inaccessible strip for one in an accessible strip, marching $\Gamma$ leftward one unit at a time. The Riemann zeta function is the second archetype. Its Dirichlet series converges only for $\Re s>1$ , but the functional equation continues it to a meromorphic function on the whole plane whose one pole sits at $s=1$ .

\zeta(s)=\sum_{n=1}^{\infty} n^{-s}\ \ (\Re s>1),\qquad \zeta(s)=2^{s}\pi^{s-1}\sin\!\Big(\tfrac{\pi s}{2}\Big)\Gamma(1-s)\,\zeta(1-s) \qquad (6)

where: $s$ — complex variable; the functional equation reflects $s\leftrightarrow 1-s$ , carrying the accessible region $\Re s>1$ onto $\Re s<0$ ; $\zeta$ has a single simple pole at $s=1$ with residue $1$ and is otherwise entire. Reading: this reflection is what assigns finite, physically meaningful values such as $\zeta(-1)=-\tfrac{1}{12}$ and $\zeta(0)=-\tfrac12$ to a series that plainly diverges there.

Derivation AUniqueness of analytic continuation

Analytic continuation is only useful if the answer is unique — if the function we build by chaining power-series disks does not secretly depend on the accidental centres and radii we chose along the way. The claim to establish is exactly this: any two holomorphic extensions of the same seed to the same connected domain must coincide identically, everywhere. The whole argument rests on the identity theorem, which converts a statement about agreement on a small overlap into a statement about agreement across the entire domain, and its proof shows precisely why holomorphy is so much more rigid than smoothness.

Step 1. Suppose $F_1$ and $F_2$ are holomorphic on a connected domain $\Omega$ and both equal the seed $f$ on some subdomain $U\subset\Omega$ . Form the difference $h:=F_1-F_2$ , which is holomorphic on $\Omega$ and vanishes identically on the open set $U$ . Why: subtracting reduces "two functions agree" to "one function vanishes", which is easier to propagate.

Step 2. Pick any point $z_0\in U$ ; it is a limit point of zeros of $h$ . If $h$ had a zero of finite order $m$ there, we could write $h(z)=(z-z_0)^m\phi(z)$ with $\phi(z_0)\neq0$ , and $h$ would be nonzero throughout a punctured neighbourhood — contradicting that $z_0$ is surrounded by zeros. Hence every derivative vanishes, $h^{(k)}(z_0)=0$ for all $k\ge0$ . Why: a nonzero holomorphic function has isolated zeros, so an accumulation of zeros forces the function to be zero to infinite order.

Step 3. With all Taylor coefficients zero, $h\equiv0$ on a whole disk about $z_0$ . Now let $A$ be the set of points of $\Omega$ where $h$ and all its derivatives vanish. $A$ is open (each such point carries a null disk) and closed (the vanishing conditions are continuous), and it is non-empty. Because $\Omega$ is connected, $A=\Omega$ , so $h\equiv0$ and $F_1\equiv F_2$ . Why: connectedness is the lever that turns a local fact into a global one — the domain cannot split into a vanishing part and a non-vanishing part.

Step 4. One caveat completes the statement: uniqueness is per-domain and per-homotopy-class. Two regions that overlap only by encircling a branch point can host genuinely different single-valued branches, because the connecting overlap is not simply connected. Why: the identity theorem needs an honest open overlap inside one connected domain; looping around a singularity does not supply one.

Go deeper · Derivation B · 1 minRe-expanding the geometric series past its disk

Derivation BRe-expanding the geometric series past its disk

We make the abstract claim that a fresh disk can poke past the old one completely concrete, on the one series everybody trusts: the geometric series $\sum_{n\ge0} z^n$ , convergent only for $|z|<1$ , whose sum is $1/(1-z)$ . Starting from that unit disk we re-expand about the interior point $z_1=-\tfrac12$ , compute the radius of the new disk directly from Cauchy–Hadamard, and show it reaches out to $z=-2$ — genuinely outside the original disk — while the pole at $z=1$ blocks any progress toward the right. This is analytic continuation in a single, fully computable step.

Step 1. Choose the new centre $z_1=-\tfrac12$ , which lies inside $|z|<1$ , so the value and all derivatives of $f$ there are available from the original series. Why: re-expansion is only legitimate from a point already inside the current disk, where the seed is known.

Step 2. Since $f(z)=1/(1-z)$ has $f^{(n)}(z)=n!/(1-z)^{n+1}$ , the new coefficients are $b_n=f^{(n)}(z_1)/n!=1/(1-z_1)^{n+1}$ . With $z_1=-\tfrac12$ this is $b_n=(2/3)^{n+1}$ . Why: the coefficients decay geometrically with ratio $2/3$ , and that ratio is exactly what sets the new radius.

Step 3. Cauchy–Hadamard gives $R_1=1/\limsup|b_n|^{1/n}=|1-z_1|=\tfrac32$ . The new disk is centred at $-\tfrac12$ with radius $\tfrac32$ , so it stretches from $z=-2$ on the left to $z=1$ on the right, touching the pole but not passing it. Why: the nearest singularity to $z_1$ is the pole at $z=1$ , at distance $\tfrac32$ , and that distance is the radius.

Step 4. The new disk contains points with $|z|>1$ — for instance $z=-1.4$ — so the continuation has escaped the original unit disk on the left. Repeating the move around the point $z=1$ eventually covers all of $\mathbb{C}\setminus\{1\}$ , reconstructing $1/(1-z)$ globally. Why: each re-expansion advances the frontier; only the genuine singularity at $z=1$ is impassable. *Note: $f^{(n)}(z)=n!/(1-z)^{n+1}$ and the Cauchy–Hadamard formula together give $R_1=|1-z_1|=3/2$ , consistent with the pole at $z=1$ .*

Go deeper · Derivation C · 1 minMultivaluedness of the logarithm around the origin

Derivation CMultivaluedness of the logarithm around the origin

The logarithm is the cleanest laboratory for multivaluedness, and it shows why a branch cut or a Riemann surface is not optional bookkeeping but a structural necessity. We begin from the principal-branch power series near $z_0=1$ , recognize $\log z$ as an antiderivative of $1/z$ , and continue it once around the origin along a circle. The loop integral of $\mathrm{d}z/z$ is a residue computation we can perform exactly; its value $2\pi i$ is precisely the jump the continued function acquires, and it is nonzero for a purely topological reason — the punctured plane is not simply connected.

Step 1. Near $z_0=1$ the principal branch is $\log z=\sum_{n\ge1}\frac{(-1)^{n+1}}{n}(z-1)^n$ , convergent for $|z-1|<1$ , and it satisfies $\dfrac{\mathrm{d}}{\mathrm{d}z}\log z=\dfrac1z$ . So $\log z$ is a continuation of an antiderivative of $1/z$ , a function holomorphic on $\mathbb{C}\setminus\{0\}$ . Why: framing $\log$ as $\int \mathrm{d}z/z$ turns its multivaluedness into a statement about a contour integral.

Step 2. Continue the antiderivative along a loop $\gamma(t)=re^{it}$ , $t\in[0,2\pi]$ , encircling the origin once counterclockwise. The continued value at the endpoint is the original value plus $\int_\gamma \mathrm{d}z/z$ . Why: chaining disks along the loop accumulates exactly the path integral of the derivative $1/z$ .

Step 3. Evaluate the increment by the residue at the simple pole: $\displaystyle\oint_{|z|=r}\frac{\mathrm{d}z}{z}=2\pi i$ . Therefore, returning to the start, $\log z\mapsto\log z+2\pi i$ ; in particular $\log 1$ moves from $0$ to $2\pi i$ . Why: the residue of $1/z$ at $0$ is $1$ , and $2\pi i$ times that residue is the loop integral.

Step 4. The origin is a logarithmic branch point, and the monodromy theorem does not apply because $\mathbb{C}\setminus\{0\}$ is not simply connected: the loop cannot be shrunk to a point without crossing the singularity. The single-valued object lives on the Riemann surface — an infinite spiral whose sheets differ by $2\pi i$ . Why: multivaluedness is the price of projecting a spiral surface down onto a flat plane.

Limiting cases

How far a continuation can go, and whether it stays single-valued, is dictated entirely by the singularities it meets on the boundary of convergence.

Case	Result	Meaning
Isolated pole	$f(z)=\dfrac{1}{1-z}=\sum_{n=0}^{\infty} z^n$	Series lives only on $\|z\|<1$ , yet the closed form continues $f$ to all of $\mathbb{C}\setminus\{1\}$ ; the boundary is breached everywhere except the single pole.
Branch point	$\log z=\log r+i\theta,\ \ z=re^{i\theta}$	Continuation around $z=0$ shifts $\theta\to\theta+2\pi$ , so $\log$ is multivalued; a branch cut or a Riemann surface is required.
Natural boundary	$f(z)=\sum_{n=0}^{\infty} z^{2^n}$	Singularities are dense on $\|z\|=1$ , so this lacunary series cannot be continued past the unit circle in any direction — no continuation exists.
Entire function	$e^z=\sum_{n=0}^{\infty}\dfrac{z^n}{n!}$	No singularities, infinite radius of convergence; one series already is the global function and continuation is trivial.

Common trapTreating a multivalued continuation as if it were single-valued

The most common and most costly error is to forget that uniqueness of analytic continuation holds only on a fixed connected domain and along a fixed homotopy class of paths. Around a branch point — $z=0$ for $\log z$ or $z^{1/2}$ , or $z=\pm i$ for $\arctan z$ — continuation along inequivalent loops returns different values, and ignoring the branch cut silently mixes sheets. The symptoms are notorious: an integral with the wrong sign, a "paradox" where $\sqrt{z}\cdot\sqrt{z}\neq z$ , or a phase off by $2\pi$ . The cure is to name the branch explicitly, track which sheet you are on, and place cuts where they belong. Rigidity guarantees a unique answer per sheet, never a unique answer across sheets.

Worked examples

1 · Continuing the Gamma function to the left (fully worked)

The integral $\Gamma(z)=\int_0^\infty t^{z-1}e^{-t}\,\mathrm{d}t$ converges only for $\Re z>0$ . Use analytic continuation to evaluate $\Gamma(-\tfrac12)$ and $\Gamma(-\tfrac32)$ , and state the residue of $\Gamma$ at its pole $z=0$ .

Step 1. The tool is the functional equation $\Gamma(z+1)=z\,\Gamma(z)$ , proved by one integration by parts of the integral for $\Re z>0$ . Rearranged as $\Gamma(z)=\Gamma(z+1)/z$ , its right-hand side is defined wherever $\Gamma(z+1)$ is, so it continues $\Gamma$ one unit to the left with each use. Iterating covers all of $\mathbb{C}$ except the non-positive integers.

Step 2. Anchor on the known value $\Gamma(\tfrac12)=\sqrt{\pi}\approx 1.772454$ . This sits in the convergent strip, so no continuation is needed to obtain it.

Step 3. Step left once: $\Gamma(-\tfrac12)=\dfrac{\Gamma(\tfrac12)}{-\tfrac12}=-2\sqrt{\pi}\approx\mathbf{-3.54491}$ — finite even though the defining integral diverges there.

Step 4. Step left again: $\Gamma(-\tfrac32)=\dfrac{\Gamma(-\tfrac12)}{-\tfrac32}=\dfrac{-2\sqrt{\pi}}{-\tfrac32}=\dfrac{4}{3}\sqrt{\pi}\approx\mathbf{2.36327}$ . The sign alternates strip by strip, a fingerprint of the poles sandwiched between them.

Step 5. Near $z=0$ , $\Gamma(z)=\Gamma(z+1)/z\approx 1/z$ since $\Gamma(1)=1$ , so $\Gamma$ has a simple pole at $z=0$ with residue $1$ , and more generally residue $(-1)^n/n!$ at $z=-n$ . Result: $\Gamma(-1/2)=-2\sqrt{\pi}\approx-3.545$ , $\Gamma(-3/2)=\tfrac43\sqrt{\pi}\approx2.363$ , $\mathrm{Res}_{z=0}\,\Gamma=1$ .

2 · The regularized value ζ(−1) = −1/12 (fully worked)

The sum $1+2+3+\cdots$ diverges, yet zeta regularization assigns it the finite value $\zeta(-1)$ . Compute $\zeta(-1)$ from the functional equation and interpret the result.

Step 1. The Dirichlet series $\zeta(s)=\sum n^{-s}$ converges only for $\Re s>1$ , so $s=-1$ is out of reach directly. Use the functional equation $\zeta(s)=2^{s}\pi^{s-1}\sin(\tfrac{\pi s}{2})\Gamma(1-s)\,\zeta(1-s)$ , which expresses $\zeta(-1)$ through $\zeta(2)$ , a convergent value.

Step 2. Evaluate each factor at $s=-1$ : $2^{-1}=\tfrac12$ ; $\pi^{-2}$ ; $\sin(-\tfrac{\pi}{2})=-1$ ; $\Gamma(2)=1$ ; and $\zeta(2)=\sum n^{-2}=\dfrac{\pi^2}{6}$ from the Basel problem.

Step 3. Multiply: $\zeta(-1)=\tfrac12\cdot\pi^{-2}\cdot(-1)\cdot 1\cdot\dfrac{\pi^2}{6}=-\dfrac{1}{2}\cdot\dfrac{1}{6}=\mathbf{-\dfrac{1}{12}}$ . The two powers of $\pi$ cancel exactly, leaving a pure rational number.

Step 4. Interpretation: $\zeta(-1)=-\tfrac1{12}$ is not a literal sum of positive integers — it is the value the analytically continued function takes at $s=-1$ . That value is what appears, through zeta regularization, in the Casimir energy between conducting plates and in the critical dimension $D=26$ of the bosonic string, where the divergent zero-point sum $\sum n$ is replaced by $\zeta(-1)$ . Result: $\zeta(-1)=-1/12$ , the finite continuation value underlying Casimir and string regularization.

Physics of the portrait

Analytic continuation is the crowning application of the idea that a holomorphic function is a single rigid object, not a patchwork of local formulas. That idea, and the constructive machinery of overlapping power-series disks, is the life's work of one mathematician who insisted analysis be built from the ground up on rigorous foundations.

Origins · 1 minKarl Weierstrass and how this was discovered

Karl Weierstrass

1815–1897 · power-series analytic continuation and the rigorous foundations of complex analysis

Karl Weierstrass came to mathematics late and by an unusual road — for fourteen years he was a schoolteacher in remote Prussian towns, doing research in isolation and publishing little, until a single 1854 paper on abelian functions brought him a doctorate and, within two years, a chair in Berlin. There he built complex analysis on the power series as the primitive object: an analytic function is a locally convergent power series together with all of its continuations. In this "Weierstrass" picture, a function element is a series with its disk of convergence, and analytic continuation is the operation of re-expanding about interior points to reach fresh disks; the totality of continuations of one element is the complete analytic function, a global object assembled from local pieces. He made rigorous the identity theorem that guarantees the process is unique, exhibited lacunary series whose circle of convergence is an impassable natural boundary, and — with his famous nowhere-differentiable function of 1872 — showed how sharply complex-analytic rigidity contrasts with the wild freedom of merely continuous real functions. Where Riemann approached the same territory geometrically, through surfaces and integrals, Weierstrass approached it algebraically, through series and estimates, and the two viewpoints together are the two lenses through which physicists still read the subject: Riemann surfaces for the topology of branch points, Weierstrass continuation for the constructive extension of $\Gamma$ , $\zeta$ , and scattering amplitudes.

History

1851

— Riemann's dissertation introduces Riemann surfaces, giving multivalued continuation a geometric home

1859

— Riemann continues the zeta function to the whole plane and states its functional equation

1876

— Weierstrass publishes his power-series theory of single-valued analytic functions and their continuation

1884

— Poincaré and Mittag-Leffler sharpen continuation and the monodromy theorem through automorphic functions

1892

— Hadamard's gap theorem pins down when a lacunary series has the unit circle as a natural boundary

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. on functions of a complex variable · Arfken & Weber, Mathematical Methods for Physicists, 7th ed., ch. 11 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 24–25 · Ahlfors, Complex Analysis, ch. 6 (analytic continuation, monodromy) · Whittaker & Watson, A Course of Modern Analysis, ch. 5. The key numerical checkpoints across this article are: $f^{(n)}(z)=n!/(1-z)^{n+1}$ , giving re-expansion radius $R_1=|1-z_1|=\tfrac32$ for $z_1=-\tfrac12$ ; $\oint \mathrm{d}z/z=2\pi i$ ; the Gamma recurrence $\Gamma(z+1)=z\Gamma(z)$ with $\Gamma(-\tfrac12)=-2\sqrt\pi\approx-3.545$ and $\Gamma(-\tfrac32)=\tfrac43\sqrt\pi\approx2.363$ ; and the functional-equation evaluation $\zeta(-1)=2^{-1}\pi^{-2}\sin(-\tfrac{\pi}{2})\Gamma(2)\zeta(2)=-\tfrac1{12}$ using $\zeta(2)=\pi^2/6$ .

Builds on: Analytic Functions and the Cauchy–Riemann Equations · Taylor Series of Analytic Functions · Zeros, Poles, and Isolated Singularities · Cauchy's Integral Formula and Derivatives. Leads to: Laurent Series · Consequences: Liouville, Maximum Modulus, Morera · Elementary Complex Functions.

Ask the tutor: Ask me to walk you through continuing $\Gamma$ or $\zeta$ into a forbidden strip step by step, or to show on the Riemann surface exactly where $\log z$ picks up its $2\pi i$ .

Does an analytic continuation always exist?

No. Whether a function element continues past its circle of convergence depends on the singularities on that circle. If they are isolated — a pole, a branch point — you can route around them and continue almost everywhere. But some functions have singularities densely packed on the entire boundary. The lacunary series $\sum z^{2^n}$ is the standard example: every root of unity of the form $e^{2\pi i p/2^q}$ is a singular point, so no overlapping disk can reach outside the unit circle in any direction. The circle is then a natural boundary, and the function has no continuation at all. Hadamard's gap theorem says any series whose nonzero exponents grow geometrically behaves this way.

Is analytic continuation really unique, or just conventionally so?

It is genuinely unique, with one precise caveat. On a fixed connected domain, the identity theorem guarantees that at most one holomorphic function extends a given seed — there is no freedom whatsoever. The caveat is topological: if you continue along a path that loops around a branch point, the endpoint value can depend on the loop's homotopy class. This is not a failure of uniqueness but a feature of the geometry — the function is single-valued on its Riemann surface, and the apparent ambiguity is only the shadow cast when you project that surface onto the plane. So the honest statement is: unique per connected domain and per homotopy class of paths, always.

Does 1 + 2 + 3 + ⋯ actually equal −1/12?

Not as an ordinary sum — the partial sums march off to infinity, and no rearrangement changes that. The precise statement is that the analytic continuation of $\zeta(s)=\sum n^{-s}$ , which equals the honest sum only for $\Re s>1$ , takes the value $-\tfrac1{12}$ at $s=-1$ . Assigning that value to " $\sum n$ " is a definition (zeta regularization), justified because it is the unique value consistent with holomorphy and because it produces the right physics: the Casimir force and the bosonic string's critical dimension both come out correct. Say " $\zeta(-1)=-\tfrac1{12}$ ", not "the sum of the integers is $-\tfrac1{12}$ ", and the paradox evaporates.

What is the difference between the identity theorem and analytic continuation?

They are two sides of one coin. Analytic continuation is the construction — the act of extending a function beyond its original domain by re-expanding series or invoking functional equations. The identity theorem is the guarantee — the statement that any such extension, if it exists, is the only one possible on a given connected domain. Without the identity theorem, continuation would be an arbitrary craft: you could imagine many different extensions and have no reason to prefer one. With it, the extension is forced, so it is meaningful to speak of the analytic continuation of $\Gamma$ or $\zeta$ . Construction supplies existence; the identity theorem supplies uniqueness.

Why do physicists trust dimensional regularization when the dimension d is an integer?

Because the loop integrals are analytic functions of $d$ , and analytic continuation extends them off the integers uniquely. One computes a Feynman integral in a range of $d$ where it converges, obtains a closed form involving $\Gamma(n-d/2)$ , and then continues that expression to $d=4-\epsilon$ . The identity theorem ensures there is only one function agreeing with the convergent integral, so the continued answer is not an arbitrary choice. Ultraviolet divergences reappear as poles in $\epsilon$ , cleanly separated from the finite physics. The scheme is trusted precisely because it preserves gauge and Lorentz symmetry — properties a momentum cutoff would spoil — and its predictions match experiment to many decimal places.

What is a Riemann surface, in one intuitive picture?

It is the natural home of a multivalued function, built so the function becomes single-valued on it. Take $\log z$ : on the plane it has infinitely many values differing by $2\pi i$ , because circling the origin adds $2\pi i$ each time. Instead of forcing all those values onto one flat plane and drawing a branch cut, imagine a spiral ramp — like a multi-storey car park with no floors, just one endlessly rising surface. Each full turn around the origin lifts you to the next level, where $\log z$ is larger by $2\pi i$ . On that spiral the logarithm is perfectly single-valued and continuous; the plane is merely the ramp's shadow, and the branch cut is where the shadow's edges overlap.

Can real-analytic functions be continued the same way?

Locally, yes — a real-analytic function on an interval is rigid in the same sense: if it vanishes on a subset with a limit point, it vanishes on the whole connected interval. The dramatic contrast is with merely smooth ( $C^\infty$ ) functions, which can vanish on an entire interval yet be nonzero elsewhere — impossible for anything analytic. The deeper payoff comes from complexifying: a real-analytic function is the restriction of a holomorphic one, and continuing it into the complex plane reveals singularities (poles, branch points) that were invisible on the real axis but that control the radius of convergence of its Taylor series.