Analytic Continuation

Why analytic continuation matters

• Riemann zeta function. The series Σ 1/n^s defines ζ on Re(s) > 1; analytic continuation extends ζ to all of ℂ except a simple pole at s = 1. The Riemann hypothesis — every nontrivial zero of ζ has Re(s) = 1/2 — is a statement about the continued ζ, not the original series. The prime number theorem, Hadamard's 1896 proof, and the Bombieri-Vinogradov theorem all depend on properties of analytically continued ζ.
• Gamma function. Γ(z) = ∫ t^{z−1} e^{−t} dt converges on Re(z) > 0. The functional equation Γ(z+1) = z Γ(z) extends Γ to all of ℂ except non-positive integers. Γ underpins the Beta function, the Stirling approximation, the digamma function, and the Wallis product — all defined via the continued Γ.
• QFT regularization. Loop integrals in quantum field theory diverge in 4 dimensions but are finite for non-integer d < 4. Compute the integral as a function of d, analytically continue d → 4 from a region of convergence; divergences appear as 1/(d−4) poles. Dimensional regularization, due to 't Hooft and Veltman, is now standard in particle physics calculations.
• L-functions and Langlands program. Dirichlet L-functions L(s, χ) = Σ χ(n)/n^s, automorphic L-functions, Hasse-Weil zeta functions for elliptic curves — all are originally defined by series convergent in a half-plane, then extended by analytic continuation. The Langlands functoriality conjectures predict these extensions agree with each other in deep ways.
• Statistical mechanics — Lee-Yang theory. The grand partition function Z(z) of a lattice gas is initially defined for real fugacity z > 0. Lee and Yang continued Z to complex z and showed that the location of zeros (the Lee-Yang circle theorem) detects phase transitions: when zero-loci accumulate on the positive real axis, a transition occurs.
• Casimir effect — physical relevance of ζ(−3). The Casimir force between parallel plates is F ∝ ζ(−3). The naive sum is divergent; the regularized value (via analytic continuation of ζ to s = −3) gives the experimentally verified force. Sphaleron decay rates and vacuum energies in cosmology share this regularization structure.
• Algorithm analysis — generating functions. Coefficient asymptotics of Σ aₙ zⁿ are extracted by analytically continuing past the radius of convergence and identifying singularities; this is the foundation of Flajolet-Sedgewick singularity analysis for combinatorial sequences.

Common misconceptions

• "ζ(−1) = −1/12 means 1 + 2 + 3 + … equals −1/12." No. The series 1 + 2 + 3 + … diverges in the standard sense — it has no finite sum. ζ(−1) = −1/12 is the value of the analytic continuation of ζ at s = −1. The function ζ is defined by the series only for Re(s) > 1; everywhere else it is defined by uniqueness of analytic continuation, which produces a different function with the same name.
• "Any function extends." No. Functions with natural boundaries (e.g., Σ z^{2ⁿ}, the modular function j on the upper half plane, lacunary Hadamard series) cannot be continued past their domain. The unit circle for Σ z^{2ⁿ} is a "wall of singularities" with no analytic extension to |z| ≥ 1.
• "Extension is path-independent." Only on simply connected domains. On punctured or multiply connected domains, continuing along different paths can give different values (monodromy). log z, √z, and Riemann surfaces of multivalued functions are governed by monodromy groups.
• "Continuation builds new analytic functions out of old ones." Not really — it identifies the unique analytic function on the larger domain whose restriction matches. The "new" function is forced to exist by uniqueness and is in some sense already there; continuation is recognition, not construction.
• "Σ converges past R" justifies continuation. No. The series Σ z^n diverges at z = 2 even though 1/(1 − z) is perfectly well-defined there. Convergence of the series is irrelevant to the continued function. The series is one of many representations of f; the function exists independently.
• "Numerical methods can compute continuations." Carefully: Padé approximants, conformal maps, and polynomial extrapolation can extend numerically, but these methods are sensitive to noise and to the location of singularities. Naive evaluation of the original power series past R fails catastrophically; specialized algorithms (chebfun's adaptive methods, Stahl's theorem on Padé convergence) handle the noise.

The identity theorem

Let f and g be analytic on a connected open set D ⊂ ℂ. If the set {z ∈ D : f(z) = g(z)} has an accumulation point in D, then f = g on all of D.

Proof sketch: let h = f − g, also analytic. At any accumulation point a ∈ D, h has a power series Σ cₙ (z − a)ⁿ on a small disc around a. Since h(zₖ) = 0 for a sequence zₖ → a, the coefficients cₙ all vanish (by induction on n, using Taylor remainder bounds). So h ≡ 0 in a neighborhood of a. The set where h ≡ 0 in a neighborhood is open (just shown) and closed (if zₖ → z* and h ≡ 0 near zₖ, then z* is an accumulation point of zeros, so h ≡ 0 near z*). By connectedness, this set is all of D, so h ≡ 0 globally.

This is a vastly stronger uniqueness than for real C^∞ functions: a smooth bump function and the zero function agree on a half-line accumulation set yet differ. Analytic = much rigid. The identity theorem is what makes the phrase "the analytic continuation" sensible — there is at most one extension to any given larger domain.

The Riemann zeta function

The series ζ(s) = Σ_{n=1}^∞ 1/n^s converges absolutely for Re(s) > 1. The integral representation

ζ(s) = (1/Γ(s)) ∫₀^∞ t^{s−1} / (e^t − 1) dt    for Re(s) > 1

extends to Re(s) > 0, s ≠ 1, by isolating the contribution of t near 0. From there, Riemann's 1859 paper presents two paths of further extension:

1. The Hankel contour deformation: integrate around a keyhole contour, picking up residues. This gives ζ(s) on all of ℂ \ {1} as a contour integral.
2. The functional equation: ζ(s) = 2^s π^{s−1} sin(πs/2) Γ(1 − s) ζ(1 − s). This relates values on Re(s) > 1/2 to values on Re(s) < 1/2; one side determines the other.

Plug in s = −1: ζ(−1) = 2^{−1} · π^{−2} · sin(−π/2) · Γ(2) · ζ(2) = (1/2)(1/π²)(−1)(1)(π²/6) = −1/12.

Plug in s = 0: ζ(0) = 2⁰ · π^{−1} · sin(0) · Γ(1) · ζ(1) — but the right side has a 0 (sin) times ζ(1) which has a pole. The product evaluates to −1/2 in the limit. Mechanism: ζ(s) − 1/(s−1) is entire; at s = 0 this gives ζ(0) − (−1) = 1 + ζ(0) such that the limiting calculation yields ζ(0) = −1/2.

Riemann then conjectured that the nontrivial zeros of ζ all have Re(s) = 1/2 — the Riemann Hypothesis, still unproven. Its truth would give the sharpest possible error term in the prime number theorem.

Geometric series, gamma, exponential

The simplest example: f(z) = Σ z^n. This series converges on |z| < 1 and diverges for |z| ≥ 1. On the disc, the sum equals 1/(1 − z). The function 1/(1 − z) is defined and analytic on ℂ \ {1}. By the identity theorem, 1/(1 − z) is the unique analytic continuation of f to any larger connected domain containing the original disc — so on ℂ \ {1}, "f" is 1/(1 − z), with a simple pole at z = 1.

Gamma function: Γ(z) = ∫₀^∞ t^{z−1} e^{−t} dt converges for Re(z) > 0. The functional equation Γ(z+1) = z Γ(z), proven by integration by parts on the integral representation, lets us extend. For Re(z) ∈ (−1, 0): define Γ(z) = Γ(z+1)/z. The right side is well-defined since Re(z+1) ∈ (0, 1) is in the original domain. This recursively extends Γ to Re(z) > −n for any n, leaving simple poles at z = 0, −1, −2, … with residues (−1)ⁿ/n! at z = −n. On all of ℂ minus non-positive integers, Γ is meromorphic.

Exponential continued from the real line: e^x for x ∈ ℝ is real-analytic with Taylor series Σ x^n/n!. The same series with z complex defines an entire function e^z, which is the analytic continuation of e^x to all of ℂ. By identity theorem, e^z is the unique entire function whose restriction to ℝ is e^x.

Monodromy and Riemann surfaces

Continuation along a path: pick a starting point a ∈ U where f is analytic, choose a path γ from a to a target point b. Cover γ with a chain of overlapping discs D₀, D₁, …, Dₙ where D₀ ⊂ U; expand f as a power series at the center of D₀, then re-expand at the center of D₁ (using the overlap with D₀), and so on. The end power series at center near b defines f(b), the continuation along γ.

If the target domain is simply connected, all paths give the same answer. If not, paths in different homotopy classes can give different answers.

Example: log z on ℂ \ {0}. Start at z = 1 with log(1) = 0. Continue along the upper half of the unit circle to z = −1: log(−1) = iπ. Continue along the lower half: log(−1) = −iπ. The two continuations differ by 2πi, which equals the change in log around one loop. The "function" log z is not single-valued on ℂ \ {0}; it lives naturally on a Riemann surface that is an infinite spiral covering of ℂ \ {0}, with log z single-valued there.

The monodromy group of a multivalued function is the group of permutations of values induced by continuation around closed loops. For log z it is ℤ (winding number); for √z it is ℤ/2; for the modular function j(τ) it is the modular group PSL(2, ℤ) acting on the upper half plane. Universal covers and Riemann surfaces are precisely the spaces on which multivalued continuations become single-valued.

Natural boundaries

Some functions cannot be analytically continued past their original domain. The classical example is the Hadamard gap series

f(z) = Σ_{n=0}^∞ z^{2ⁿ} = z + z² + z⁴ + z⁸ + z¹⁶ + …

This converges on |z| < 1. Hadamard proved that every point on |z| = 1 is a singularity — f cannot be continued past any arc of the unit circle. The unit circle is a "natural boundary" for f.

The mechanism: at z = e^{2πi p/2^k} (a 2^k-th root of unity), the partial sum f_N(z) → ∞ as N → ∞, since infinitely many terms z^{2ⁿ} with n ≥ k equal 1. These dyadic points are dense in the unit circle, forcing every point to be singular by continuity. Generalizations (Ostrowski's gap theorem) say lacunary series — power series Σ a_k z^{n_k} with n_{k+1}/n_k bounded away from 1 — have natural boundaries on their disc of convergence.

The modular function j(τ) on the upper half plane has the real line as a natural boundary; the boundary is dense in the real line modulo the action of SL(2, ℤ). Modular forms, Eisenstein series, theta functions — all live on the upper half plane with a natural boundary on the real axis, providing rich structure but no possible extension across.

Methods of continuation

• Power series re-expansion. Standard textbook procedure. Given f on disc D₀ with center a₀, re-expand at a point a₁ ∈ D₀ near the boundary; the new disc D₁ may extend past D₀. Repeat. Computationally feasible for analytic functions known to high precision.
• Functional equations. The most economical: if f satisfies f(z) = something_in_terms_of_f(other_argument), the equation extends f to wherever the other argument lies in the original domain. Used for ζ, Γ, Bernoulli polynomials, Hurwitz zeta, modular forms.
• Integral representations. Replace a power series by an integral that has a wider region of convergence. Hankel contour for ζ, Mellin-Barnes integrals for hypergeometric functions, Cauchy integral formula for analytic functions on contours.
• Borel summation. Recover divergent series via the Borel sum: f(z) ≈ Σ aₙ zⁿ, B(z) = Σ aₙ zⁿ/n! (which often converges past R), then f(z) = ∫₀^∞ B(zt) e^{−t} dt. Gives analytic continuation past the original disc; foundation of resurgence theory.
• Padé approximants. Rational approximations [m/n] (z) = P_m(z)/Q_n(z) matched to the Taylor series; their poles approximate the singularities of f, and the rational extrapolates past R. Numerically robust if singularity structure is mild.
• Conformal mapping methods. Map the original domain to a disc, expand f as a power series there, then map back. For domains with complicated boundaries, this gives accurate numerical continuation; Driscoll's Schwarz-Christoffel toolbox is a standard implementation.

Where analytic continuation shows up

• Particle physics — dimensional regularization. Loop integrals in d = 4 spacetime dimensions diverge; in d = 4 − 2ε they converge for small ε > 0. Compute the integral as a meromorphic function of d, expand around d = 4, isolate the 1/ε pole, subtract via renormalization. Standard procedure since 't Hooft and Veltman, used in every QED, QCD, and standard model calculation.
• Number theory — Riemann hypothesis. The behavior of analytically continued ζ(s) controls prime distribution. RH says all nontrivial zeros lie on Re(s) = 1/2; assuming RH, the prime counting function π(x) = li(x) + O(√x log x) — the optimal error. Trillions of zeros computed satisfy RH; a proof is the Clay Millennium Problem.
• Cosmology — Casimir energy. The vacuum energy between two plates in QED is ⟨E⟩ ∝ Σ ωₙ; the sum diverges. ζ-function regularization defines ⟨E⟩ = (1/2) ζ(−1) · (ω-scale) = −1/24 · (scale), giving the experimentally verified attractive Casimir force. Vacuum energies in compact extra dimensions follow the same template.
• Statistical physics — Lee-Yang theorem. The grand partition function Z(z) of a lattice gas is a polynomial in fugacity z; Lee and Yang showed its zeros lie on |z| = 1 for ferromagnetic Ising models. Phase transitions correspond to zeros accumulating on the positive real axis as system size grows. Z(z) extended to complex z is the analytic continuation; its zero locus is the physics.
• Combinatorics — singularity analysis. Generating functions F(z) = Σ aₙ zⁿ are continued past R; aₙ ∼ (singular term) · (radius)^{−n} · n^{−α} · log^β n with exponents α, β extracted from the dominant singularity of F. Flajolet-Sedgewick, "Analytic Combinatorics" (2009), is the canonical reference.
• Quantum mechanics — resonances. The Hamiltonian H of a metastable state has real eigenvalues for bound states. Resonances correspond to poles of the resolvent (z − H)⁻¹ continued past the real axis to a second sheet of a Riemann surface; the imaginary parts give decay rates. Complex scaling methods (Aguilar-Balslev-Combes) compute these continuations numerically.