Spherical Bessel and Neumann Functions

Every function family that survives in physics earns its keep twice: once as a solution and once as a toolkit. The previous article established that free space, sorted by angular momentum, answers in spherical Bessel functions — $j_\ell$ regular at the origin, $n_\ell$ singular there. This article opens the toolkit: where the family comes from, what each member looks like, and the five working facts — closed forms, small-argument laws, large-argument phases, recurrences, and the Wronskian — that every later matching computation will draw on. The organizing surprise is that the family is elementary. Unlike their cylindrical cousins $J_\nu$ , which genuinely need infinite series, every spherical Bessel function is a finite combination of sines, cosines, and powers of $1/x$ : $j_0=\sin x/x$ , $n_0=-\cos x/x$ , and Rayleigh's differentiation formula manufactures all the rest. The reason is a dimensional accident with real content — the substitution $R=Z/\sqrt{x}$ maps the spherical equation onto the cylindrical one at order $\ell+\tfrac12$ , and half-integer Bessel functions are the ones that collapse to elementary form. Three dimensions sits exactly half a dimension away from two, and the half is what makes it solvable in closed form. Two behaviors carry all the physics. Near the origin, $j_\ell\sim x^\ell/(2\ell+1)!!$ — the centrifugal turnstile that priced ℓ-channels out of the previous article's thermal-neutron problem. Far away, $j_\ell\to\sin(x-\ell\pi/2)/x$ : every channel oscillates with the same envelope, distinguished only by a phase offset of a quarter turn per unit of ℓ — the flat stage onto which the scattering chapter will walk its phase shifts. Everything printed here — every coefficient, zero, recurrence, and the $-\ell\pi/2$ law itself — is machine-verified against symbolic reduction and scipy's implementations; the well spectra those verified zeros quantize belong to the spherical-well article two serials ahead.

Key takeaways

Definitions: $j_\ell$ = regular solution of the spherical Bessel equation $x^2f''+2xf'+[x^2-\ell(\ell+1)]f=0$ ; $n_\ell$ = the singular partner (also "Neumann" or $y_\ell$ ).
The half-integer bridge: $j_\ell(x)=\sqrt{\pi/2x}\,J_{\ell+1/2}(x)$ — verified symbolically for $\ell=0,1,2$ ; half-integer order is why the family is elementary.
Rayleigh's formulas: $j_\ell=(-x)^\ell\big(\tfrac1x\tfrac{d}{dx}\big)^\ell\tfrac{\sin x}{x}$ , $n_\ell=(-x)^\ell\big(\tfrac1x\tfrac{d}{dx}\big)^\ell\big(-\tfrac{\cos x}{x}\big)$ — one seed each, differentiate to any ℓ (verified against closed forms and scipy).
Small argument: $j_\ell\to x^\ell/(2\ell+1)!!$ , $n_\ell\to-(2\ell-1)!!/x^{\ell+1}$ — double factorials, verified for $\ell\le3$ ; the $(kr)^\ell$ suppression and the origin eviction in one line each.
Large argument: $j_\ell\to\sin(x-\ell\pi/2)/x$ , $n_\ell\to-\cos(x-\ell\pi/2)/x$ — same $1/x$ envelope, quarter-turn phase per ℓ; good to 10% beyond $x\approx\ell(\ell+1)/0.2$ (measured: 8.4, 28.6, 58.3, 97.6 for $\ell=1\ldots4$ ).
Machinery: three-term recurrence $f_{\ell-1}+f_{\ell+1}=(2\ell+1)f_\ell/x$ (swept $\ell\le5$ ); Wronskian $j_\ell n_\ell'-n_\ell j_\ell'=1/x^2$ (symbolic $\ell\le2$ , numeric $\ell\le5$ ) — the independence certificate matching problems cash.
Zeros (root-found): $j_0$ : $\pi,2\pi$ ; $j_1$ : $4.4934,7.7253$ ; $j_2$ : $5.7635,9.0950$ ; $j_3$ : $6.9879$ — the raw material of Particle in a Spherical Well and Infinite Spherical Box.
Key idea: one elementary family, two limiting laws — $x^\ell$ at the origin, phase-lagged sines at infinity — supplies the entire radial vocabulary of free 3D quantum mechanics.

The core insight

Spherical Bessel functions are the half-integer-order Bessel family in elementary disguise: Rayleigh's derivative formula generates every $j_\ell,n_\ell$ from $\sin x/x$ and $-\cos x/x$ , with $x^\ell/(2\ell+1)!!$ growth at the origin, $\sin(x-\ell\pi/2)/x$ oscillation at infinity, and a $1/x^2$ Wronskian certifying the pair's independence — the complete radial toolkit for $V=0$ and for every matching problem beyond.

Where this lives today

Any code that touches waves in three dimensions ships a spherical Bessel module: partial-wave scattering engines in nuclear and atomic physics, electromagnetic Mie-scattering solvers (droplets, aerosols, plasmonic nanoparticles), and the fast multipole methods that make million-atom electrostatics tractable all evaluate $j_\ell,n_\ell$ by the recurrences verified here — with the downward-recursion discipline of the warning box baked into the libraries. Cosmology is an unexpectedly heavy user: the spherical harmonic transform of the cosmic microwave background radial-projects every Fourier mode through $j_\ell(k\chi)$ , so the acoustic peaks in the CMB power spectrum are read through exactly these oscillation patterns. And ultrasound tomography and room acoustics still solve Helmholtz problems in the same alphabet Rayleigh standardized — the family's third century of employment.

The physical picture

Watch one function, say $j_2$ , from left to right. Near the origin it hugs zero — the $x^2$ suppression of the centrifugal turnstile, invisible until $x\sim\ell$ . Around $x\approx\ell$ it wakes up (classically: the radius where a particle of angular momentum $\ell\hbar$ and wavenumber $k$ can first exist, $kr\approx\ell$ ), climbs to its first and largest maximum, and then settles into steady ripples that weaken as $1/x$ . Three regimes — asleep, waking, rippling — and the two asymptotic laws of this article are the sleeping and rippling ends made quantitative.

Figure

Three regimes of j_ℓ: centrifugal sleep, classical turning, free ripples

The Neumann functions are the mirror family: where $j_\ell$ sleeps politely at the origin, $n_\ell$ diverges as $-(2\ell-1)!!/x^{\ell+1}$ — the previous article evicted them from whole-space problems for exactly this crime. But the far field cannot tell the families apart except by phase: $n_\ell$ ripples with the same $1/x$ envelope, a quarter turn out of step. That near/far asymmetry is the family's division of labor — origin behavior decides admissibility, far-field phase carries physics — and it is why the evicted $n_\ell$ returns to full employment the moment a problem has a region excluding the origin (the next serial's Hankel combinations, the well matchings after that).

From the radial equation to a known family

The free radial equation of the previous article, written for the full radial factor $R(x)$ with $x=kr$ , is the spherical Bessel equation:

x^2\frac{d^2R}{dx^2}+2x\frac{dR}{dx}+\big[x^2-\ell(\ell+1)\big]R=0 \qquad (1)

where: the $2x$ in the first-derivative term (rather than the cylindrical $x$ ) is the fingerprint of three dimensions — it is the same $2R'/r$ that the $u=rR$ trick of Effective Potential and the Centrifugal Barrier exists to remove. Equation (1) is second-order, so its solution space is two-dimensional for each ℓ; this article's whole business is choosing, naming, and characterizing a good basis pair for that space.

Derivation AHalf a dimension from the cylinder: the reduction to order ℓ plus one-half

We connect (1) to the classical Bessel equation and extract the deep reason spherical Bessel functions are elementary. The strategy is a single substitution chosen to kill the first-derivative mismatch between the spherical ( $2x$ ) and cylindrical ( $x$ ) equations — the same weight-transfer logic as $u=rR$ , applied halfway.

Step 1. Substitute $R(x)=Z(x)/\sqrt{x}$ . The derivatives reshuffle ( $R'=Z'/\sqrt x-\tfrac12Z x^{-3/2}$ , and similarly for $R''$ ); collecting terms multiplies out to

x^2Z''+xZ'+\Big[x^2-\big(\ell+\tfrac12\big)^2\Big]Z=0 \qquad (2)

— exactly Bessel's equation of order $\nu=\ell+\tfrac12$ . The identity $\ell(\ell+1)+\tfrac14=(\ell+\tfrac12)^2$ does the bookkeeping (the same $\tfrac14$ that appeared in the fall-to-center discriminant of Effective Potential and the Centrifugal Barrier — recurring because both are half-line weight transfers). Why "half a dimension": the radial Laplacian in $d$ dimensions carries $(d-1)/x$ in its first-derivative term, and the weight substitution $R=Z\,x^{-(d-2)/2}$ always lands on Bessel's equation of order $\ell+(d-2)/2$ . Two dimensions gives integer order (the genuinely non-elementary $J_m$ ); three gives half-integer. The spherical family sits precisely between integer orders — and half-integer Bessel functions terminate into elementary form.

Step 2. So $j_\ell(x)=\sqrt{\pi/2x}\,J_{\ell+1/2}(x)$ (the $\sqrt{\pi/2}$ is convention, chosen to make $j_0=\sin x/x$ exactly), and $n_\ell(x)=\sqrt{\pi/2x}\,Y_{\ell+1/2}(x)$ — equivalently $(-1)^{\ell+1}\sqrt{\pi/2x}\,J_{-\ell-1/2}(x)$ , since at half-integer order the two signs of $\nu$ are already independent. Verified symbolically: the bridge identity reproduces the closed forms below for $\ell=0,1,2$ .

Step 3. Why elementarity follows: $J_{1/2}(x)=\sqrt{2/\pi x}\,\sin x$ (one line from the series — the half-integer gamma functions telescope), and Bessel's recurrence machinery then generates every higher half-integer order by differentiation and division — operations that keep sines, cosines, and powers closed among themselves. No new transcendental function is ever created. The efficient packaging of that machinery is Rayleigh's formula, derived next. Verified: substitution (2) symbolic; bridge identity $j_\ell=\sqrt{\pi/2x}J_{\ell+1/2}$ for $\ell=0,1,2$ exact (article verifier).

Go deeper · Derivation B · 2 minRayleigh's formula: the whole family from two seeds

Derivation BRayleigh's formula: the whole family from two seeds

We derive the generating formula $j_\ell=(-x)^\ell\big(\tfrac1x\tfrac{d}{dx}\big)^\ell\,\tfrac{\sin x}{x}$ and its Neumann mirror, then harvest the closed forms and both asymptotic laws from it. The strategy: verify the operator $\hat{D}=\tfrac1x\tfrac{d}{dx}$ steps ℓ up by one, then induct.

Step 1. The stepping claim: if $f_\ell$ solves (1) at ℓ, then $g=-x\,\hat{D}f_\ell\cdot$ (bookkeeping) — precisely, $f_{\ell+1}=(-x)^{\ell+1}\hat{D}^{\ell+1}f_0$ solves (1) at $\ell+1$ . Rather than a page of operator algebra, establish the equivalent derivative relation directly: $f_\ell'=f_{\ell-1}-\tfrac{\ell+1}{x}f_\ell$ holds for the family (verified numerically across $\ell\le4$ ), and reading it backwards shows $\hat{D}$ acting on $f_\ell/x^\ell$ produces $-f_{\ell+1}/x^{\ell+1}$ — the compact statement $\hat{D}\big(x^{-\ell}f_\ell\big)=-x^{-(\ell+1)}f_{\ell+1}$ , which iterates from $\ell=0$ into Rayleigh's formula.

Step 2. Harvest closed forms by turning the crank once and twice ( $j_0=\sin x/x$ seed; $n$ -family seed $n_0=-\cos x/x$ ):

j_1=\frac{\sin x}{x^2}-\frac{\cos x}{x},\qquad j_2=\Big(\frac{3}{x^3}-\frac1x\Big)\sin x-\frac{3\cos x}{x^2} \qquad (3)

with mirrors $n_1=-\cos x/x^2-\sin x/x$ and $n_2=(-3/x^3+1/x)\cos x-3\sin x/x^2$ . Each of the six satisfies (1) — verified symbolically, and the Rayleigh crank reproduces all of them plus $j_3$ , $n_2$ against scipy on a grid.

Step 3. Small-argument law. Feed the sine series through the crank: each application of $\hat{D}$ eats the constant term and divides the leading power's coefficient by successive odd numbers; after ℓ turns, $j_\ell\to x^\ell/(2\ell+1)!!$ with $(2\ell+1)!!=1\cdot3\cdot5\cdots(2\ell+1)$ (verified for $\ell\le3$ : coefficients $1,\tfrac13,\tfrac1{15},\tfrac1{105}$ ). The cosine seed's $1/x$ leads to $n_\ell\to-(2\ell-1)!!/x^{\ell+1}$ (verified $\ell\le2$ ). Physical read: the double factorial is the accumulated centrifugal toll — the $(kr)^\ell$ suppression priced channel by channel in the previous article now has its exact constant.

Step 4. Large-argument law and the quarter-turn phase. At large $x$ , the dominant term of each $\hat{D}$ application is the one differentiating the trig factor (every division by $x$ costs a power; differentiation of $\sin$ costs none). Each derivative turns $\sin(x)\to\cos(x)\to-\sin(x)\dots$ — a phase retreat of $\pi/2$ per step — while the prefactor keeps exactly one $1/x$ :

j_\ell(x)\xrightarrow{x\gg\ell}\frac{\sin(x-\ell\pi/2)}{x},\qquad n_\ell(x)\to-\frac{\cos(x-\ell\pi/2)}{x} \qquad (4)

Why the phase is the physics: the far field of every channel is the same wave, offset by $\ell\pi/2$ . Any potential's entire far-field effect will be one more offset $\delta_\ell$ added to this baseline — the scattering chapter's phase shift has its zero point defined by equation (4). Onset honesty, measured not guessed: the first correction is $\ell(\ell+1)/2x$ relative to the $1/x$ envelope (near the sine's zeros the local relative error is correspondingly larger), so (4) holds to envelope-tolerance $\varepsilon$ only beyond $x\approx\ell(\ell+1)/2\varepsilon$ — at 10%, measured onsets $8.4,\,28.6,\,58.3,\,97.6$ for $\ell=1,2,3,4$ against predictions $10,30,60,100$ (computed). The asymptote is not a fixed- $x$ statement; it is a statement about $x$ relative to $\ell^2$ . Verified: derivative relation swept; Rayleigh crank == closed forms == scipy; double-factorial constants exact; far-field phases at $x=50$ for $\ell\le4$ ; onset table (article verifier).

Two working identities complete the kit. The three-term recurrence — one relation running the whole ladder in either direction:

f_{\ell-1}(x)+f_{\ell+1}(x)=\frac{2\ell+1}{x}\,f_\ell(x),\qquad f\in\{j,n\} \qquad (5)

where: this is how software actually evaluates the family (with the numerical caveat that upward recursion is unstable for $j_\ell$ once $\ell>x$ — the honest fine print of the small-argument law: errors pick up the $n_\ell$ -like growing contamination; libraries recurse downward instead, a lesson verified painfully by every generation of computational physicists). Swept numerically for $\ell\le5$ , both families. And the Wronskian — the independence certificate:

j_\ell(x)\,n_\ell'(x)-n_\ell(x)\,j_\ell'(x)=\frac{1}{x^2} \qquad (6)

where: a nonvanishing Wronskian says the pair spans the full two-dimensional solution space at every $x$ — so any solution of (1) anywhere is $A\,j_\ell+B\,n_\ell$ , the one-line justification behind every interior/exterior matching the next two serials perform. The $1/x^2$ (rather than a constant) is the standard weight for an equation whose first-derivative coefficient is $2/x$ ; verified symbolically for $\ell\le2$ and numerically for $\ell\le5$ .

Figure

The zero ladder: root-found zeros of j_ℓ interlace — the raw material of spherical-well spectra

Limiting cases

case	result	meaning
$x\to0$ , $j$ family	$j_\ell\to x^\ell/(2\ell+1)!!$	centrifugal sleep; exact toll constants $1,\tfrac13,\tfrac1{15},\tfrac1{105}$
$x\to0$ , $n$ family	$n_\ell\to-(2\ell-1)!!/x^{\ell+1}$	eviction from whole-space problems; hire back for exteriors
$x\gg\ell(\ell+1)/2\varepsilon$	$j_\ell\to\sin(x-\ell\pi/2)/x$ to tolerance $\varepsilon$	flat stage + quarter-turn phase per ℓ; onset measured, not folklore
$\ell=0$	$j_0=\sin x/x$ , $n_0=-\cos x/x$	the seeds; everything else is Rayleigh's crank
$\ell>x$ upward recursion	numerically unstable	growing-solution contamination — recurse downward
$d$ dimensions	order $\ell+(d-2)/2$	elementary only in odd $d$ ; 3D's luck is half-integer order

Common trapTwo naming conventions and one lethal recursion direction

Bookkeeping hazards that produce real wrong answers. First, names: the singular family appears as $n_\ell$ (most quantum texts), $y_\ell$ (Abramowitz & Stegun, scipy's `spherical_yn`) — same functions; but cylindrical $Y_\nu$ and Neumann's own normalization conventions differ across older tables, so when importing formulas match against the closed forms (3), not against symbols. Second, the double factorials: $(2\ell+1)!!\neq(2\ell+1)!$ — for $\ell=2$ that is $15$ versus $120$ , a factor-8 error in a small-argument amplitude if confused. Third, computation: evaluating $j_\ell(x)$ for $\ell\gtrsim x$ by upward recursion (5) amplifies rounding error catastrophically because the recurrence's other solution ( $n$ -like) grows; the standard cure is downward recursion from a high trial ℓ with final rescaling. If your hand-rolled Bessel routine disagrees with scipy at $\ell=10$ , $x=2$ , this is why.

Worked examples

Example 1 — run the crank and audit it three ways. Generate $j_2$ from scratch and certify it.

Step 1. Rayleigh, twice: $\hat{D}\tfrac{\sin x}{x}=\tfrac{x\cos x-\sin x}{x^3}$ ; apply $\hat{D}$ again and multiply by $(-x)^2$ : $j_2=\big(\tfrac{3}{x^3}-\tfrac1x\big)\sin x-\tfrac{3}{x^2}\cos x$ (computed symbolically).

Step 2. Audit one — the ODE: substitute into (1) with $\ell=2$ ; the residual vanishes identically (computed). Audit two — the bridge: $\sqrt{\pi/2x}\,J_{5/2}(x)$ expands to the same expression (computed). Audit three — the limits: series-expand: leading term $x^2/15=x^2/(2\cdot2+1)!!$ ✔; at $x=50$ , $j_2(50)=0.004083$ versus the asymptote $\sin(50-\pi)/50=0.005247$ — a deviation of $0.06$ of the envelope $1/50$ , exactly the predicted correction scale $\ell(\ell+1)/2x=6\%$ (it looks like 22% of the local value only because $x=50$ sits near a node — the envelope, not the local value, is the honest yardstick; both numbers computed).

Step 3. Moral: three independent certificates — differential equation, half-integer bridge, limiting laws — and the function is earned, not copied. Answer: $j_2$ as in (3), triple-audited.

Example 2 — how big is an ℓ = 1 wave at a nuclear surface? Put numbers on the sleep regime for a 1 MeV neutron at $r=3$ fm.

Step 1. Wavenumber: $k=\sqrt{2m_nE}/\hbar$ with $E=1$ MeV gives $k=0.219\,\text{fm}^{-1}$ , so $x=kr=0.66$ .

Step 2. Sleep-law estimate: $j_1(0.66)\approx x/3=0.219$ ; exact value $0.2103$ (the law is good to 4% here — the first correction $-x^3/30=-0.0095$ closes the gap to 0.1%, computed). Compare $j_0(0.66)=0.929$ : the p-wave amplitude is $4.4\times$ smaller, its probability $19.5\times$ smaller — noticeable but not negligible at 1 MeV, which is why fast-neutron physics keeps a few partial waves while thermal physics (previous article: $x\sim10^{-4}$ ) keeps one.

Step 3. The $n$ -family contrast at the same point: $n_1(0.66)=-2.75$ (computed from the closed form) — an order of magnitude larger than $j_1$ and growing inward. In an exterior matching problem this size mismatch is legitimate physics (both members serve); in a whole-space problem it is the eviction notice, quantified. Answer: $j_1=0.210$ vs $j_0=0.929$ at $x=0.66$ — the turnstile at work, with the small-argument law good to 4% and repaired to 0.1% by its first correction.

Origins · 1 minFriedrich Wilhelm Bessel and how this was discovered

Friedrich Wilhelm Bessel

1784–1846 · Königsberg Observatory

The functions were born doing celestial bookkeeping, twenty years before anyone waved at them in a laboratory. Bessel — whose route from merchant's apprentice to Prussia's great astronomer, and whose 1828 inequality auditing quantum expansions, are told in Complete Orthonormal Bases — hit the family head-on in his analytical work on planetary perturbation (memoirs of 1817–1824): expanding the elliptic-motion problem of Kepler orbits in trigonometric series, the expansion coefficients refused to be anything simpler than the integrals now called $J_\nu$ . He tabulated them, derived their recurrences and their differential equation, and used them to weigh the mutual tugs of planets — special functions as industrial astronomy. The spherical cousins have a braided pedigree: their elementary closed forms circulate from Euler's and Poisson's wave computations onward, Rayleigh's Theory of Sound made the $\hat{D}$ -crank formula standard equipment, and the singular family picked up Carl Neumann's name after his 1867 treatise systematized the second solution (the same Neumann whose boundary-value work anchors an electrostatics article of the E&M track — different hat). Nomenclature, as usual, rewards the systematizer as much as the discoverer. Bessel himself would have appreciated the article's zero tables most: his life's method was to trust nothing he had not recomputed to the last digit — the working philosophy this article's verifier merely automates.

History

1732

— Daniel Bernoulli meets Bessel-type series in the hanging-chain problem

1817

— Bessel begins the systematic theory: planetary-perturbation memoirs introduce

J_\nu

and their recurrences

1824

— Bessel's definitive memoir: the differential equation, tabulation, asymptotics

1867

— Carl Neumann's treatise systematizes the second solution; the

n_\ell

naming follows

1877

— Rayleigh's Theory of Sound deploys the derivative formula and the spherical family for acoustics

1927

— Quantum scattering adopts the family wholesale as the radial alphabet of partial waves

Going further

Griffiths & Schroeter §4.1.3 — the family's quantum introduction and first use.
Abramowitz & Stegun §10.1 (or DLMF Ch. 10) — the canonical reference tables behind every formula here.
Sakurai & Napolitano §6.4 — the family deployed in scattering.
Watson, A Treatise on the Theory of Bessel Functions — the complete mathematical monograph, for the curious.

Provenance note: A verifier should recompute: (1) the reduction of the free radial equation to the spherical Bessel equation and the indicial roots $s=\ell+1,-\ell$ (Griffiths §4.1.3); (2) the small- and large-argument asymptotics against Abramowitz & Stegun §10.1; (3) the $\ell=0,1$ closed forms and the recurrence relation; (4) the infinite-well spectrum $E_{n0}=\hbar^2n^2\pi^2/2ma^2$ and the first $\ell=1$ zero $\beta_{11}\approx4.4934$ ; (5) the Rayleigh plane-wave expansion coefficients $(2\ell+1)i^\ell$ . — Items (1)–(3) are executed in this article's verify script (ODE residuals for six closed forms; the half-integer bridge; Rayleigh-crank cross-checks against scipy; double-factorial constants; far-field phases with the measured 10%-onset table; recurrence and Wronskian sweeps), with the $u$ -form indicial roots being The Radial Equation's verified property, cited. Item (4): the zeros $\pi$ , $4.4934$ , $5.7635$ , $6.9879$ , $7.7253$ , $9.0950$ are root-found in this article's verifier, but the spectra they quantize are owned by Particle in a Spherical Well and Infinite Spherical Box per chapter ownership — the base's infinite-well derivation is RELOCATED there (noted in both provenances) and its energy formula is verified in that article's script. Item (5) belongs to Partial Wave Expansion of Plane Waves, likewise by ownership. Edition note: the small-/large-argument asymptotics derivation relocated here FROM the free-particle article's base (its former second derivation) now lives in this article's Rayleigh derivation, Steps 3–4.

Are n_ℓ ever needed if they're always evicted at the origin?

Constantly. Eviction applies only when $r=0$ belongs to the physical domain. Every matching problem with an interior region — a spherical well, a hard core, a nucleus inside an atom — has an exterior where both families are legal and the physics chooses the mix: the finite spherical well matches interior $j_\ell$ to an exterior combination, and scattering states far from any target are $j_\ell\cos\delta_\ell-n_\ell\sin\delta_\ell$ . The singular family is half of every answer that isn't free space.

Why do the closed forms terminate while cylindrical Bessel functions don't?

The series for $J_\nu$ involves $\Gamma(k+\nu+1)$ in denominators. At half-integer $\nu$ these gamma functions pair up with the $\sqrt{x}$ prefactor so that the whole series resums into $\sin$ and $\cos$ — a consequence of $\Gamma(\tfrac12)=\sqrt{\pi}$ propagating through the recursion. At integer $\nu$ no such resummation exists; $J_0$ is genuinely a new transcendental function. Equivalently: the 1D half-line problem (which 3D reduces to via $u=rR$ ) has elementary solutions, and $j_\ell$ inherits them; the 2D radial problem reduces to nothing simpler.

Is there a generating function like the one for Legendre polynomials?

Yes, and it is physics in disguise: $\frac{\sin\sqrt{x^2-2xt}}{\sqrt{x^2-2xt}}=\sum_\ell \frac{t^\ell}{\ell!}\,\text{(Rayleigh-type terms)}$ — but the version actually used daily is the plane-wave expansion itself, $e^{ikz}=\sum(2\ell+1)i^\ell j_\ell(kr)P_\ell(\cos\theta)$ , which packages the whole family against the whole Legendre family. That identity is the next-but-one article's centerpiece; treating it as "the generating function of $j_\ell$ " is a legitimate reading.

How do the zeros interlace, and does it matter?

Between consecutive zeros of $j_\ell$ lies exactly one zero of $j_{\ell+1}$ (and vice versa) — Sturm-type interlacing inherited from the oscillation theory of The Node Theorem's grammar. Practically it means the infinite-well level ordering ( $1s,1p,1d,2s,\dots$ by increasing zero) can never produce two levels of adjacent ℓ crossing as you scan well size: the ordering by zeros in this article's table is rigid. It also underwrites root-finding: brackets for the $(n{+}1)$ -th zero come free from the $n$ -th.

What exactly goes wrong with upward recursion numerically?

The recurrence (5) is shared by both families. Computing $j_\ell$ upward, rounding error injects a tiny admixture of the $n$ -family — and for $x<\ell$ that admixture grows like $(2\ell-1)!!/x^{\ell+1}$ while your target shrinks like $x^\ell/(2\ell+1)!!$ . By $\ell\approx x+10$ the contamination dominates entirely. Downward recursion reverses the roles: the unwanted solution decays, errors are laundered out, and one final normalization against $j_0$ fixes the scale. This is a textbook instance of a general law: recurse in the direction where your solution is the growing one.

Do I ever need non-integer ℓ?

In this chapter, never — ℓ is quantized by the angular problem. But the machinery permits any order, and two neighbors use it: the Coulomb problem's long tail effectively shifts ℓ to non-integer values (Coulomb wave functions carry a continuous order parameter), and dimensional regularization tricks in scattering theory analytically continue in ℓ (Regge theory, one honesty line). If you meet $j_{\ell+\epsilon}$ , the bridge formula to $J_{\ell+1/2+\epsilon}$ still holds — only the elementarity is lost.