Special Functions from Equations and Boundary Data

Special functions recur because separation of variables, nondimensionalization, and a rescaling of the dependent variable reduce many linear field equations to a small set of canonical second-order ordinary differential equations. The canonical equation fixes a two-dimensional local solution space. It does not, by itself, fix the physical solution: regularity, decay, radiation or positive-frequency data select a direction in that space, while a Wronskian, flux, or inner product fixes the remaining normalization.

This distinction is the practical answer to why the same Bessel, Airy, Hankel, and hypergeometric functions appear in radial, curved, thermal, and bounded problems. The useful object is not merely a function name but a package: equation, domain, parameters, branches, endpoint data, continuation path, and normalization. The de Sitter mode calculation below makes every part of that package explicit.

Required background. Linear ODEs, Evolution Operators, and Wronskians. It supplies two-dimensional solution spaces, fundamental bases, boundary maps, connection matrices, and Wronskian normalization. This page specializes that machinery to equations with distinguished regular and singular points.

Solution spaces, branches, and normalization

Basic use of the functions below does not require WKB, steepest descent, or Mellin methods.

For two functions of one variable,

$$W_z[f,g]=f(z)g'(z)-f'(z)g(z).$$

Unless a different branch is stated, complex powers use the principal logarithm. An asymptotic formula is asserted only on the ray or in the sector that accompanies it.

The equation, the basis, and the selected solution

Consider

$$y''+P(z)y'+Q(z)y=0$$

on a simply connected domain that avoids any singularities and branch cuts. At an ordinary point, $P$ and $Q$ are analytic and both solutions have Taylor series. An isolated singular point $z_0$ is regular singular when $(z-z_0)P(z)$ and $(z-z_0)^2Q(z)$ extend analytically to $z_0$; an isolated singular point that fails this test is irregular. Frobenius exponents describe the leading local powers at a regular singular point. Integer differences between those exponents are resonant and can replace a naive second power-series solution by a logarithmic one.

Three logically separate choices are easy to conflate.

1. Canonical reduction. A change of variable and a prefactor map the original equation to a standard one. The prefactor, transformed domain, and derivative Jacobian remain part of the answer.
2. Basis choice. One selects a pair adapted to a location: origin, infinity, a turning point, or another singular point. Different pairs are related by connection matrices.
3. Physical selection and normalization. Boundary, initial, analyticity, or radiation data select a linear combination. A source strength, Wronskian, conserved flux, or canonical inner product then fixes its scale.

Two endpoint conditions need not select a solution for every parameter. In a boundary-value problem they can instead give a determinant condition whose zeros quantize an eigenvalue. At a singular endpoint, even “regular” may be insufficient: square-integrability and the domain of the operator can require additional analysis.

The recurring reductions can be summarized compactly.

Canonical structure	Useful local basis	Typical selecting datum
Radial equation with an inverse-square term	$J_\nu,Y_\nu$	Origin behavior or a finite boundary
Oscillatory Bessel equation at infinity	$H_\nu^{(1)},H_\nu^{(2)}$	Incoming/outgoing phase and flux
Sign-reversed radial equation	$I_\nu,K_\nu$	Regularity or exponential decay on a stated ray
Linear normal form $y''-zy=0$	$\operatorname{Ai},\operatorname{Bi}$	Recessive behavior in a stated sector
Three regular singular points	Local ${}_2F_1$ bases	Analyticity and connection data at $0,1,\infty$

This equation-first view of Bessel, modified Bessel, Airy, and hypergeometric functions, together with sample integral representations, is introduced in Costin 2016, graduate course notes, §1.2, printed p. 9, PDF. Transformations and representations for Airy, modified Bessel, and hypergeometric functions are developed there in §2.4, printed pp. 75–82. The singular-point classification used here is summarized in NIST DLMF 2026, §2.7(i).

Bessel functions: an origin-adapted basis

The Bessel equation is

$$z^2y''+zy'+(z^2-\nu^2)y=0.$$

It has a regular singular point at $z=0$, with exponents $\pm\nu$, and an irregular singular point at infinity. On the principal branch, a standard solution is

$$J_\nu(z) = \left(\frac z2\right)^\nu \sum_{j=0}^{\infty} \frac{(-z^2/4)^j}{j!\,\Gamma(j+\nu+1)}.$$

For noninteger $\nu$, an independent solution can be written

$$Y_\nu(z) = \frac{\cos(\pi\nu)J_\nu(z)-J_{-\nu}(z)}{\sin(\pi\nu)}.$$

The apparent singularity at integer order is resolved by taking a limit. It cannot be repaired by using $J_{-n}$ instead, because $J_{-n}=(-1)^nJ_n$ for $n\in\mathbb Z$. The resulting $Y_n$ supplies the missing logarithmic or singular solution.

For real $z\to0^+$ and real $\nu>0$,

$$J_\nu(z) \sim \frac{(z/2)^\nu}{\Gamma(\nu+1)}, \qquad Y_\nu(z) \sim -\frac{\Gamma(\nu)}{\pi} \left(\frac 2z\right)^\nu.$$

Thus, when the domain includes the origin and ordinary finiteness is the correct boundary condition, $J_\nu$ is the adapted choice. If a regular solution also obeys a Dirichlet condition at $z=kR$, the allowed $k$ may be the zeros of $J_\nu(kR)$. A punctured domain, a point source, or a different operator domain can instead require a $Y_\nu$ contribution. The basis is independent because

$$W_z[J_\nu,Y_\nu]=\frac{2}{\pi z}.$$

The same logic appears after a sign change:

$$z^2y''+zy'-(z^2+\nu^2)y=0.$$

For real order $\nu\ge0$ and positive real $z$, $I_\nu(z)$ is finite at the origin and grows exponentially as $z\to+\infty$, whereas $K_\nu(z)$ decays at infinity but is singular at the origin. Which one is physical depends on which endpoint or source condition the problem actually imposes. A radial QFT example in which near-source boundary data fix the ratio of two Bessel solutions is worked out in Burgess 2021, § 13.3.2, pp. 359–360.

Definitions, branches, limiting forms, and Wronskians are collected in NIST DLMF 2026, § 10.2, NIST DLMF 2026, § 10.5, and NIST DLMF 2026, § 10.7.

Hankel functions: a wave-adapted basis

Hankel functions are not solutions of a new equation. They are the combinations

$$H_\nu^{(1)}=J_\nu+iY_\nu, \qquad H_\nu^{(2)}=J_\nu-iY_\nu,$$

chosen because their large-positive-$z$ phases separate traveling waves. For fixed $\nu$ and $z\to+\infty$,

$$H_\nu^{(1)}(z) \sim \sqrt{\frac{2}{\pi z}} e^{i(z-\pi\nu/2-\pi/4)},$$ $$H_\nu^{(2)}(z) \sim \sqrt{\frac{2}{\pi z}} e^{-i(z-\pi\nu/2-\pi/4)}.$$

Their Wronskian is

$$W_z[H_\nu^{(1)},H_\nu^{(2)}] = -\frac{4i}{\pi z}.$$

For a radial field with time dependence $e^{-i\omega t}$ and positive radial argument $z=kr$, $H^{(1)}$ is outgoing and $H^{(2)}$ is incoming. Reversing the time convention reverses that interpretation. Likewise, a time-dependent mode can have $z=-k\eta$, so the sign relating $z$ to time must be applied before either superscript is called positive frequency.

These formulas use the principal square root and the principal Hankel branches, with their standard cut along the negative real axis. Continuation around the origin changes the sheet and can mix the displayed basis. The large-argument phases and their sectors are given in NIST DLMF 2026, §10.17(i); the Wronskian is NIST DLMF 2026, §10.5.

Airy functions: boundary data on a stated ray

The Airy equation,

$$y''-zy=0,$$

has the entire solutions $\operatorname{Ai}(z)$ and $\operatorname{Bi}(z)$, with

$$W_z[\operatorname{Ai},\operatorname{Bi}]=\frac1\pi.$$

For $x\to+\infty$ on the positive real axis and $\xi=2x^{3/2}/3$,

$$\operatorname{Ai}(x) \sim \frac{e^{-\xi}}{2\sqrt{\pi}\,x^{1/4}}, \qquad \operatorname{Bi}(x) \sim \frac{e^{\xi}}{\sqrt{\pi}\,x^{1/4}}.$$

Decay on that ray therefore selects $\operatorname{Ai}$. On the negative real axis both standard solutions oscillate:

$$\operatorname{Ai}(-x) \sim \frac{\sin(\xi+\pi/4)}{\sqrt{\pi}\,x^{1/4}}, \qquad \operatorname{Bi}(-x) \sim \frac{\cos(\xi+\pi/4)}{\sqrt{\pi}\,x^{1/4}}.$$

It is therefore false without qualification that “$\operatorname{Ai}$ decays at infinity.” In the complex plane, different rotated Airy solutions are recessive in different sectors. The fractional power in the asymptotic description requires a branch, even though the Airy functions themselves are entire.

A differential equation near a simple linear zero often reduces locally to the Airy equation. That structural fact explains its recurrence; the detailed approximation and sector-to-sector matching are developed in WKB and Eikonal Methods and Turning-Point Matching. Exact Airy equations, Wronskians, connections, and sectorial asymptotics are tabulated in NIST DLMF 2026, §9.2 and NIST DLMF 2026, §9.7.

Hypergeometric functions: connection data between singular points

The Gauss hypergeometric equation is

$$z(1-z)y''+ [c-(a+b+1)z]y'-aby=0.$$

Its three singular points $0,1,\infty$ are regular, with exponent pairs

$$\begin{array}{c|c} z=0 & 0,\ 1-c \\ z=1 & 0,\ c-a-b \\ z=\infty & a,\ b \end{array}$$

where the exponents at infinity mean behaviors proportional to $z^{-a}$ and $z^{-b}$ after choosing branches. A second-order Fuchsian equation with exactly three regular singular points can be mapped to this form by a Möbius change of variable and a prefactor. That universality is why ${}_2F_1$ appears in so many separated equations.

For $|z|<1$ and $c\notin\{0,-1,-2,\ldots\}$,

$${}_2F_1(a,b;c;z) = \sum_{n=0}^{\infty} \frac{(a)_n(b)_n}{(c)_n\,n!}z^n$$

is the solution normalized to one at $z=0$. When $c\notin\mathbb Z$, a generic local basis there is

$$f_1(z)={}_2F_1(a,b;c;z),$$ $$f_2(z)= z^{1-c} {}_2F_1(a-c+1,b-c+1;2-c;z),$$

with

$$W_z[f_1,f_2] = (1-c)z^{-c}(1-z)^{c-a-b-1}.$$

Suppose $\delta=c-a-b$ and the parameters are generic, with $\delta\notin\mathbb Z$ and no displayed gamma factor at a pole. On principal branches in the cut domain

$$|\arg z|<\pi, \qquad |\arg(1-z)|<\pi,$$

define

$$C_0= \frac{\Gamma(c)\Gamma(\delta)} {\Gamma(c-a)\Gamma(c-b)}, \qquad C_1= \frac{\Gamma(c)\Gamma(-\delta)} {\Gamma(a)\Gamma(b)}.$$

Then continuation from the basis at zero to one gives

$$\begin{aligned} {}_2F_1(a,b;c;z) ={}&C_0\,{}_2F_1(a,b;1-\delta;1-z)\\ &+C_1(1-z)^\delta {}_2F_1(c-a,c-b;1+\delta;1-z). \end{aligned}$$

The gamma ratios are entries of a connection matrix, not decorative constants. They tell how the solution normalized at one singular point decomposes into a basis adapted to another. If an exponent difference is an integer, the generic basis or formula must be replaced by a parameter limit; logarithmic solutions and cancellations between divergent gamma factors can appear. Directly evaluating the generic formula at resonance is both mathematically wrong and numerically ill-conditioned.

The equation, exponent pairs, local bases, Wronskians, and connection formulas are collected in NIST DLMF 2026, §15.10. DLMF also uses a bold Olver-normalized hypergeometric function in some transformation formulas; it differs from ordinary ${}_2F_1$ by a gamma normalization, so notation must be checked before coefficients are copied. A de Sitter propagator example in which the field equation becomes hypergeometric and analyticity plus an $i\epsilon$ prescription select the solution appears in Burgess, Introduction to Effective Field Theory, Exercise 10.6, printed pp. 271–272.

Controlled QFT example: a normalized de Sitter mode

Consider a minimally coupled real scalar of mass $m$ in the spatially flat patch of four-dimensional de Sitter spacetime,

$$\mathrm ds^2 = a(\eta)^2(\mathrm d\eta^2-\mathrm d\mathbf x^2), \qquad a(\eta)=-\frac1{H\eta}, \qquad \eta<0.$$

With the Fourier convention

$$\phi(\eta,\mathbf x) = \int\frac{\mathrm d^3k}{(2\pi)^3} \phi_{\mathbf k}(\eta)e^{i\mathbf k\cdot\mathbf x},$$

the quadratic action is

$$S = \frac12\int\mathrm d\eta\,\mathrm d^3x\,a^2 \left[ (\phi')^2-(\boldsymbol\nabla\phi)^2-a^2m^2\phi^2 \right].$$

Set $v_{\mathbf k}=a\phi_{\mathbf k}$. For $k=|\mathbf k|>0$,

$$v_k''+ \left(k^2+a^2m^2-\frac{a''}{a}\right)v_k=0,$$

or

$$v_k''+ \left[ k^2-\frac{\nu^2-1/4}{\eta^2} \right]v_k=0, \qquad \nu^2=\frac94-\frac{m^2}{H^2}.$$

Restrict this example to $0\le m^2<9H^2/4$ and choose $\nu=\sqrt{9/4-m^2/H^2}>0$. Put $z=-k\eta>0$. Direct substitution gives the two-dimensional solution space

$$v_k(\eta) = \sqrt{-\eta} \left[ A_kH_\nu^{(1)}(z)+B_kH_\nu^{(2)}(z) \right].$$

Impose the early-time positive-frequency condition

$$v_k(\eta) \longrightarrow \frac{e^{-ik\eta}}{\sqrt{2k}} \qquad (-k\eta\to+\infty).$$

Because $e^{-ik\eta}=e^{+iz}$ in this patch, the large-$z$ phase selects $H_\nu^{(1)}$, not $H_\nu^{(2)}$. Matching its amplitude and constant phase gives

$$v_k(\eta) = \frac{\sqrt{-\pi\eta}}{2} \exp\!\left[\frac{i\pi(2\nu+1)}4\right] H_\nu^{(1)}(-k\eta).$$

The asymptotic formula immediately returns $e^{-ik\eta}/\sqrt{2k}$. Normalization supplies an independent check. For real $\nu$ and positive $z$, $(H_\nu^{(1)})^*=H_\nu^{(2)}$, so

$$\begin{aligned} W_\eta[v_k,v_k^*] &= \frac\pi4(-\eta)(-k) W_z[H_\nu^{(1)},H_\nu^{(2)}]\\ &= \frac\pi4(-\eta)(-k) \left(-\frac{4i}{\pi z}\right) =i. \end{aligned}$$

Equivalently, $i(v_k^*v_k'-v_k^{*\prime}v_k)=1$, the canonical commutator normalization. The prefactor, the Jacobian $\mathrm dz/\mathrm d\eta=-k$, and the argument sign are all needed for this check.

For $m=0$, $\nu=3/2$ and the half-integer Hankel function reduces to

$$v_k(\eta) = \frac{e^{-ik\eta}}{\sqrt{2k}} \left(1-\frac{i}{k\eta}\right).$$

It satisfies the mode equation, the Wronskian condition, and the early-time limit directly. This is a patchwise $k>0$ mode calculation for a prescribed linear equation. It does not by itself establish a globally regular vacuum, resolve the massless zero-mode infrared problem, compute particle production, or define a renormalized observable. The canonical normalization and positive-frequency selection follow Baumann 2012, § 12.2, pp. 56–58 and Problem 7, p. 66. Baumann uses the $(-+++)$ signature; deriving the action above with the site’s $(+---)$ convention gives the same canonically normalized mode equation and Wronskian condition.

Developed comparisons of solvable mode evolution, Bogoliubov data, and particle-production benchmarks continue in Parametric-Oscillator and Solvable Production Benchmarks.

A reproducible selection workflow

Given a separated second-order mode equation, use this sequence.

1. Nondimensionalize and map. Record the original domain, dimensions, Fourier and time phases, the new variable, any dependent-variable prefactor, and the transformed endpoints.
2. Classify singular points. Find ordinary, regular singular, and irregular points. Compute local exponents and flag integer differences or coalescing singularities.
3. Choose an independent local basis. Adapt it to the relevant endpoint, ray, or sector. State every branch and cut; use a resonant limiting basis when the generic pair degenerates.
4. Impose the actual data. Apply regularity, decay, boundary, source, incoming/outgoing, positive-frequency, or analyticity conditions as linear equations on the coefficients.
5. Transport when necessary. Use a connection matrix along a specified path. Check its determinant against the ratio of the two Wronskians.
6. Normalize. Fix the remaining scale with a source jump, unit flux, canonical Wronskian, Klein–Gordon inner product, or another declared convention.
7. Validate independently. Substitute into the original equation, check endpoint asymptotics in the actual sector, verify the Wronskian or current, and test a simplifying parameter value.

The output is a selected solution or mode together with its domain, branch, normalization, connection coefficients, and, in a boundary-value problem, possibly an eigenvalue condition. The symbolic cost of evaluating a named function is often small. The difficult work is the reduction, continuation, and conditioning: connection coefficients can contain large cancellations, and a basis well-conditioned at one endpoint can be unusable at another.

Stop and reformulate when the proposed boundary data do not define a unique or self-adjoint problem, a continuation path crosses a cut without a prescription, singularities coalesce and change the canonical equation, or a coupled or nonlinear system cannot be reduced to the claimed scalar form. Also stop before assigning a state or observable when gauge constraints, renormalization, or global spacetime input is still missing.

Common pitfalls

Treating a function name as boundary data. Saying that an equation is “Bessel” identifies a solution space, not a solution. State the domain and the condition that selects the coefficient ratio.

Equating regularity with admissibility. A finite local solution need not be square-integrable or belong to the required self-adjoint domain. Singular endpoints must be checked with the problem’s measure and operator domain.

Using $J_\nu$ and $J_{-\nu}$ at integer order. They become linearly dependent when $\nu\in\mathbb Z$. Use the limiting $Y_n$ solution or another verified independent basis.

Assigning frequency from a Hankel superscript. The assignment depends on the time phase and on how the Hankel argument depends on time. In the de Sitter example, $z=-k\eta$ is the sign that makes $H^{(1)}$ positive frequency.

Calling $\operatorname{Ai}$ globally decaying. Its behavior depends on the complex ray or sector. The function is entire; the branches belong to its asymptotic fractional powers and to the continuation convention.

Using a generic hypergeometric connection formula at resonance. Integer exponent differences can introduce logarithms, while separate divergent gamma factors may cancel only after a parameter limit is taken.

Accepting the software’s default sheet. A library can evaluate the principal branch accurately while returning the wrong continuation for the physical problem. Test the ODE residual, branch-side limits, asymptotics, and Wronskian.

Confusing exact modes with physical observables. A normalized closed-form mode does not alone fix a global state, a detector response, a particle number, or a renormalized expectation value.

Check your understanding

1. Retrieval: match basis to data

Match each datum to the most natural listed solution: (a) finiteness at a radial origin for real $\nu>0$; (b) an outgoing positive-$r$ wave with time factor $e^{-i\omega t}$; (c) decay as $x\to+\infty$ for the Airy equation; (d) analyticity and value one at $z=0$ for the hypergeometric equation.

Solution

The matches are (a) $J_\nu$, provided ordinary finiteness is the correct endpoint condition; (b) $H_\nu^{(1)}$, for positive real argument and the stated time convention; (c) $\operatorname{Ai}$ on the positive real axis; and (d) ${}_2F_1(a,b;c;z)$ when $c\notin\{0,-1,-2,\ldots\}$. Each answer includes the condition under which the label is valid.

2. Counterexample: a false Bessel basis

Why do $J_n(z)$ and $J_{-n}(z)$ fail to form a fundamental basis when $n\in\mathbb Z$? What replaces the second member?

Solution

For integer order,

$$J_{-n}(z)=(-1)^nJ_n(z),$$

so their Wronskian vanishes and they span only one direction. The independent solution is $Y_n$, defined by the $\nu\to n$ limit of the noninteger-order combination. Near zero it supplies the singular or logarithmic behavior that the repeated $J$ solution lacks.

3. Calculation: exponents at the hypergeometric origin

Insert $y=z^\rho(1+O(z))$ into the Gauss hypergeometric equation and find the two indicial exponents at $z=0$. When should a logarithmic second solution be expected?

Solution

Keeping the lowest power $z^{\rho-1}$ gives

$$\rho(\rho-1)+c\rho = \rho(\rho+c-1)=0.$$

Hence $\rho=0$ and $\rho=1-c$. Their difference is $1-c$. When this difference is an integer, equivalently $c\in\mathbb Z$, the generic Frobenius basis can degenerate and a logarithmic solution can appear. The precise resonant form must be obtained by recurrence or a parameter limit.

4. QFT transfer: select and normalize the mode

For $z=-k\eta>0$, start from $v_k=\sqrt{-\eta}(A H_\nu^{(1)}(z)+B H_\nu^{(2)}(z))$. Which coefficient vanishes under the early-time condition $v_k\to e^{-ik\eta}/\sqrt{2k}$, and what magnitude must the remaining coefficient have? Explain why this does not yet compute particle production.

Solution

Because $e^{-ik\eta}=e^{iz}$, the $H^{(1)}$ phase is selected and $B=0$. Using

$$H_\nu^{(1)}(z) \sim \sqrt{\frac{2}{\pi z}} e^{i(z-\pi\nu/2-\pi/4)}$$

shows that

$$A= \frac{\sqrt\pi}{2} e^{i\pi(2\nu+1)/4}$$

The stated early-time equality fixes this phase; if one imposed only the positive-frequency ray, a constant phase could instead be absorbed into the annihilation operator. Its magnitude is $|A|=\sqrt\pi/2$. The Hankel Wronskian then gives $W_\eta[v_k,v_k^*]=i$. Particle production would require a specified second mode basis or detector, a comparison prescription, and the associated Bogoliubov or response calculation; none follows from one normalized mode alone.

Synthesis and continuations

Bessel, Airy, Hankel, and hypergeometric functions recur because they are canonical bases for common singularity patterns of second-order linear ODEs. The equation determines the local space; endpoint, causal, and analytic data select a combination; and a conserved normalization fixes its scale. A result is reproducible only when the variable map, prefactor, domain, branches, connection path, and normalization accompany the function name.

Mellin Transforms and Scaling Asymptotics provides a different route from transform singularities to powers and logarithms. Heat Kernels, Zeta Functions, and Spectral Determinants next organizes spectral information rather than individual mode solutions. Developed time-dependent production physics continues at the curved-spacetime benchmark linked above.

References

• Daniel Baumann, TASI Lectures on Inflation, arXiv:0907.5424v2 (2012), §12.2, printed pp. 56–58, and Problem 7, printed p. 66. Canonical mode normalization, positive-frequency data, and the Hankel solution.

• C. P. Burgess, Introduction to Effective Field Theory, Cambridge University Press, first published 2021, Exercise 10.6, printed pp. 271–272, and §13.3.2, printed pp. 359–360. Hypergeometric de Sitter propagation and radial Bessel solutions selected by analytic and boundary data.

• Ovidiu Costin, Applied Complex Variables and Asymptotics II: Course Notes, PDF, Ohio State University, 2016, §1.2, printed p. 9, and §2.4, printed pp. 75–82. Equation-first introduction, transformations, integral representations, and connection behavior for the recurring families.

• NIST Digital Library of Mathematical Functions, version 1.2.7, released June 15, 2026, National Institute of Standards and Technology, §2.7(i), “Regular Singularities: Fuchs–Frobenius Theory”, Chapter 9, “Airy and Related Functions”, Chapter 10, “Bessel Functions”, and §15.10, “Hypergeometric Differential Equation”. Canonical equations, branches, bases, Wronskians, connection formulas, and sectorial asymptotics.