Modified Bessel Functions

Every time the Laplace, Helmholtz, or diffusion equation is separated in cylindrical coordinates, a radial ordinary differential equation is left behind — and its character is decided by a single sign. If the axial separation constant makes the radial motion oscillatory, the equation is Bessel's and its solutions are the ringing $J_\nu$ and $Y_\nu$ . If instead the sign flips so the radial term reads $-x^2$ rather than $+x^2$ , the equation is the modified Bessel equation and its solutions no longer oscillate: they grow or decay exponentially. Those solutions, the modified Bessel functions $I_\nu(x)$ and $K_\nu(x)$ , are to non-propagating radial fields what $e^{\pm x}$ are to a one-dimensional exponential problem. The cleanest way to see them is by analytic continuation. Rotate the argument by ninety degrees in the complex plane, $x \mapsto ix$ , and $+x^2$ becomes $-x^2$ : the modified functions are Bessel functions of imaginary argument. The regular solution $I_\nu(x) = i^{-\nu}J_\nu(ix)$ is built to be real and finite at the origin; the second solution $K_\nu(x)$ — the Macdonald function — is the combination engineered to decay at infinity, the continuation of the outgoing Hankel function $H^{(1)}_\nu$ . This pairing mirrors, term for term, the relationship between $\{J_\nu, Y_\nu\}$ and the travelling-wave Hankel functions. Physically $K_\nu$ is the screened tail — the radial Green's function of the modified Helmholtz operator $-\nabla^2+\kappa^2$ , the kernel of Yukawa and Debye screening, the field in a fibre cladding. $I_\nu$ is the finite interior partner kept when a field must stay regular on the axis. Everything you know about Bessel functions — series, recurrences, a constant Wronskian, integral representations, generating functions — carries straight over, with factors of $i$ tracked.

Key takeaways

The modified Bessel functions $I_\nu(x)$ and $K_\nu(x)$ are the two independent solutions of the modified Bessel equation $x^2 y'' + x y' - (x^2+\nu^2)y = 0$ , got from Bessel's equation by the rotation $x \to ix$ .
$I_\nu$ (first kind) is regular at the origin and grows as $e^x/\sqrt{2\pi x}$ ; $K_\nu$ (second kind, the Macdonald function) diverges at the origin and decays as $\sqrt{\pi/2x}\,e^{-x}$ .
Formally $I_\nu(x) = i^{-\nu} J_\nu(ix)$ and $K_\nu(x) = \tfrac{\pi}{2} i^{\nu+1} H_\nu^{(1)}(ix)$ — Bessel functions evaluated on the imaginary axis, hence "modified".
They are the natural radial functions with no oscillation: screened (Yukawa/Debye) potentials, steady diffusion and heat flow, and cylindrical fields below cutoff.
Both obey three-term recurrences and the exact Wronskian $W[K_\nu,I_\nu]=1/x$ , which fixes the normalisation of $K_\nu$ and underlies every Green's-function construction.
Key idea: flip the sign of the $x^2$ term in Bessel's equation — equivalently rotate the argument by $90^\circ$ — and the oscillatory pair $J_\nu, Y_\nu$ becomes the exponentially growing and decaying pair $I_\nu, K_\nu$ ; all the Bessel machinery survives the rotation with tracked factors of $i$ .

The core insight

Flip the sign of the $x^2$ term in Bessel's equation — equivalently rotate the argument by $90^\circ$ in the complex plane, $x\to ix$ . The oscillatory solutions $J_\nu, Y_\nu$ become the exponentially growing and decaying solutions $I_\nu, K_\nu$ ; every structural fact (series, recurrences, Wronskian, asymptotics) carries over with carefully tracked factors of $i$ .

Where this lives today

The Macdonald function is the quiet workhorse of screening. In plasma and electrolyte physics the Debye–Hückel potential of a line charge is $K_0(\kappa r)$ and its point-charge cousin is the Yukawa form $e^{-\kappa r}/r$ — the object Yukawa borrowed in 1935 to give the nuclear force a range. In lattice field theory the free massive propagator in position space is a modified Bessel function, so $K_\nu$ is what a Monte-Carlo two-point function is fitted against to extract a particle mass, and the same $K_0$ is the Ornstein–Zernike large-distance form of a correlation near a critical point. Photonics leans on the $J_\nu$ -core, $K_\nu$ -cladding matching to design single-mode fibres, with cutoff $V<2.405$ . And outside physics, $I_0$ normalises the von Mises distribution on the circle and the Rice distribution of signal-plus-noise amplitude, which governs magnitude images in MRI and fading in wireless links.

The physical picture

The one-dimensional templates say everything. The equation $y''+k^2 y=0$ gives oscillations $\sin kx,\cos kx$ ; the equation $y''-\kappa^2 y=0$ gives exponentials $e^{\pm\kappa x}$ . Bessel's equation is the cylindrical version of the first; the modified Bessel equation is the cylindrical version of the second. So $I_\nu$ plays the growing exponential and $K_\nu$ the decaying one, each dressed with the $x^{-1/2}$ falloff of radial spreading in two dimensions. The first figure puts the two worlds side by side: the ringing $J_0$ with its endless zeros, and the monotone $I_0$ , $K_0$ that replace it once the sign is flipped.

Figure

Bessel's equation gives the oscillatory J₀(x) with infinitely many zeros at x ≈ 2.405, 5.520, 8.654 …; flipping the sign of the x² term rotates the argument by x → ix and replaces oscillation with the monotone growing I₀(x) and the decaying K₀(x), which at x = 1 take the values I₀ = 1.2661 and K₀ = 0.4210

The rotation $x\mapsto ix$ is more than bookkeeping: at infinity it is exactly what converts oscillation into exponential behaviour, an outgoing wave $e^{ix}$ becoming a decaying tail $e^{-x}$ . Two consequences dominate the applications. First, $K_\nu$ carries a singularity at the origin — logarithmic for $\nu=0$ , a power law $x^{-\nu}$ for $\nu>0$ — that encodes a point source, while its exponential decay encodes screening over a length $1/\kappa$ . Second, because every term of the $I_\nu$ series is positive, $I_\nu$ has no real zeros: it is a monotone, node-free radial profile, the opposite of the endlessly oscillating $J_\nu$ . The second figure makes the exponential asymptotics literal, plotting both functions on a logarithmic axis where they straighten into lines whose product collapses to the exact Wronskian.

Figure

On a logarithmic vertical axis the modified Bessel functions become near-straight exponentials — I₀ ∼ eˣ/√(2πx) climbs while K₀ ∼ √(π/2x)·e⁻ˣ falls — so their product I₀K₀ → 1/(2x); the exact Wronskian x·W = x(K₀I₀′ − K₀′I₀) = 1 is the algebraic residue left after the exponentials cancel

The core equations

Everything below is one differential equation and its two solutions, read from several angles. The starting point is the modified Bessel equation — Bessel's equation with the sign of the $x^2$ term reversed.

x^2\,\frac{\mathrm{d}^2 y}{\mathrm{d}x^2} + x\,\frac{\mathrm{d}y}{\mathrm{d}x} - \left(x^2 + \nu^2\right) y = 0 \qquad (1)

where: $x$ — real, usually positive radial argument; $\nu$ — order, any complex number, most often integer or half-integer; $y$ — the solution, general form $A\,I_\nu(x)+B\,K_\nu(x)$ . The lone minus sign in front of $x^2$ is the entire difference from Bessel's equation, and it is what turns oscillation into exponential behaviour. The regular solution is a power series that differs from the $J_\nu$ series only by the loss of the alternating sign. Because every term is then positive for $x>0$ , the function climbs monotonically and never crosses zero.

I_\nu(x) = \sum_{k=0}^{\infty} \frac{1}{k!\,\Gamma(k+\nu+1)} \left(\frac{x}{2}\right)^{2k+\nu} \qquad (2)

where: $k$ — summation index; $\Gamma$ — the gamma function, generalising the factorial $(k+\nu)!$ to non-integer order; $(x/2)^{2k+\nu}$ — the positive monomial that guarantees monotone growth and no real zeros. The leading term $\propto x^\nu$ makes $I_\nu$ regular at the origin. That lost sign is not magic — it is the factor $i^{2k}=(-1)^k$ picked up on rotation. Precisely, $I_\nu$ is a Bessel function of imaginary argument, with a phase factor that keeps it real.

I_\nu(x) = i^{-\nu} J_\nu(ix), \qquad J_\nu(x) = \sum_{k=0}^\infty \frac{(-1)^k}{k!\,\Gamma(k+\nu+1)}\left(\frac{x}{2}\right)^{2k+\nu} \qquad (3)

where: $J_\nu$ — Bessel function of the first kind, whose alternating sign is flipped to $+1$ by $i^{2k}$ ; $i^{-\nu}$ — the phase factor that renders $I_\nu$ real. Read the two series together and the only change is $(-1)^k\to +1$ . The second, independent solution must decay at infinity. Rather than a bare series, $K_\nu$ is the antisymmetric combination of $I_{\pm\nu}$ that kills the growing exponential and leaves only the decaying tail — the Macdonald function.

K_\nu(x) = \frac{\pi}{2}\,\frac{I_{-\nu}(x) - I_\nu(x)}{\sin(\nu\pi)} \qquad (4)

where: $I_{-\nu}$ — the negative-order series, independent of $I_\nu$ when $\nu$ is not an integer; $\sin(\nu\pi)$ — vanishes at integer $\nu$ , forcing an L'Hôpital limit; $\pi/2$ — the normalisation that fixes the asymptotic $K_\nu\sim\sqrt{\pi/2x}\,e^{-x}$ and the Wronskian. At integer $n$ , $I_{-n}=I_n$ and the ratio is $0/0$ , resolved below. The two solutions are tied together by a Wronskian that is not merely constant but exactly $1/x$ , independent of the order — the workhorse of every Green's-function computation.

W[K_\nu, I_\nu](x) = K_\nu(x) I_\nu'(x) - K_\nu'(x) I_\nu(x) = \frac{1}{x} \qquad (5)

where: $W$ — the Wronskian of the two solutions; $I_\nu',K_\nu'$ — derivatives with respect to $x$ . Nonzero everywhere, it confirms independence; its precise value $1/x$ is what the $\pi/2$ in $K_\nu$ was chosen to deliver. The modified functions climb and descend a ladder in $\nu$ by three-term recurrences. The one sharp difference from the ordinary case is a sign: the $2\nu/x$ relation is a difference of neighbours, not a sum.

I_{\nu-1}(x) - I_{\nu+1}(x) = \frac{2\nu}{x}\, I_\nu(x), \qquad I_{\nu-1}(x) + I_{\nu+1}(x) = 2\, I_\nu'(x) \qquad (6)

where: $\nu\pm 1$ — neighbouring orders linked by the recurrence. Contrast the ordinary Bessel sum $J_{\nu-1}+J_{\nu+1}=\tfrac{2\nu}{x}J_\nu$ — the swap from sum to difference is a direct fingerprint of the $i$ rotation. For integer order the whole family is packed into one generating function, the modified twin of the Jacobi–Anger expansion.

e^{\frac{x}{2}\left(t + \frac{1}{t}\right)} = \sum_{n=-\infty}^{\infty} I_n(x)\, t^{n} \qquad (7)

where: $t$ — a nonzero expansion variable; $I_n$ — the integer-order functions as Laurent coefficients. Setting $t=e^{i\theta}$ gives $e^{x\cos\theta}=\sum_n I_n(x)e^{in\theta}$ , the source of the $I_0$ normalisation in circular statistics. Finally, $K_\nu$ has a clean integral representation that makes its exponential decay manifest.

K_\nu(x) = \int_0^\infty e^{-x\cosh t}\cosh(\nu t)\,\mathrm{d}t, \qquad x>0 \qquad (8)

where: $\cosh t\ge 1$ — forces exponential decay in $x$ ; $\cosh(\nu t)$ — the even kernel encoding the order. Because the exponent is most negative near $t=0$ , a saddle-point estimate returns $K_\nu\sim\sqrt{\pi/2x}\,e^{-x}$ in one line.

Derivation AThe modified Bessel equation as a rotation of Bessel's equation

The goal is to show that the modified Bessel equation is not a new equation at all but Bessel's own equation viewed along the imaginary axis: the substitution $z=ix$ flips the sign of the highest algebraic term and turns oscillation into exponential behaviour. We carry the chain rule through with every factor of $i$ tracked, so the identity $I_\nu(x)=i^{-\nu}J_\nu(ix)$ emerges with its phase fixed by reality rather than inserted by hand.

Step 1. Begin with Bessel's equation in the complex variable $z$ , whose regular solution is $J_\nu(z)$ : $z^2 w'' + z w' + (z^2-\nu^2)w = 0$ . Why: this is the object we will rotate; its $+z^2$ is the term whose sign we intend to invert.

Step 2. Substitute $z=ix$ , so that $\mathrm{d}/\mathrm{d}z=-i\,\mathrm{d}/\mathrm{d}x$ and $\mathrm{d}^2/\mathrm{d}z^2=-\mathrm{d}^2/\mathrm{d}x^2$ , while $z^2=-x^2$ . Why: rotating the argument by ninety degrees carries $+z^2$ into $-x^2$ ; the derivative factors of $-i$ conspire to leave the first two terms unchanged.

Step 3. Collect the pieces: $z^2 w''\to x^2 y''$ , $zw'\to (ix)(-i)y'=xy'$ , and $(z^2-\nu^2)w\to(-x^2-\nu^2)y$ , giving $x^2 y'' + x y' - (x^2+\nu^2)y=0$ . Why: only the sign of the $x^2$ term changes — this is equation (1), the modified Bessel equation, obtained with no new physics.

Step 4. Fix the phase of the solution. The series for $J_\nu(ix)$ carries $(-1)^k(i)^{2k+\nu}=i^\nu$ overall, so multiplying by $i^{-\nu}$ cancels that phase and leaves the all-positive series (2); hence $I_\nu(x)=i^{-\nu}J_\nu(ix)$ is real. Why: this pins down the definition of the first-kind modified function as the real, origin-regular solution, and shows the alternating sign of $J_\nu$ was converted to $+1$ by exactly the rotation. *Why this works term by term: the factor $(-1)^k i^{2k}=1$ for every $k$ , so the alternating signs of the $J_\nu$ series are exactly cancelled by the rotation, reproducing the all-positive series (2) from series (3); numerically $I_0(1)=1.26607$ equals $J_0(i)$ with the phase $i^{-\nu}$ removed.*

Go deeper · Derivation B · 1 minConstancy and value of the Wronskian

Derivation BConstancy and value of the Wronskian

The goal is to establish that the Wronskian of $K_\nu$ and $I_\nu$ is not just proportional to $1/x$ but equal to it exactly, with the constant fixed independently of order. Proportionality follows from Abel's identity applied to the self-adjoint form; the constant is then nailed down by the small-argument leading terms and the gamma-function recursion, which makes the two contributions add to unity.

Step 1. Write equation (1) in self-adjoint form $(xy')' - (x+\nu^2/x)y=0$ ; the coefficient of $y'$ in standard form is $1/x$ , so Abel's identity gives $\tfrac{\mathrm{d}}{\mathrm{d}x}(xW)=0$ and therefore $W=C/x$ . Why: Abel's theorem determines the shape of the Wronskian from the first-derivative coefficient alone, with no knowledge of the solutions themselves.

Step 2. Insert the small- $x$ leading behaviour for $\nu>0$ : $I_\nu\sim (x/2)^\nu/\Gamma(\nu+1)$ and $K_\nu\sim\tfrac12\Gamma(\nu)(2/x)^\nu$ , and differentiate each. Why: the constant $C$ is the same for all $x$ , so it can be read off in whatever limit is easiest, and the origin is where the leading terms are simplest.

Step 3. Form $xW=x(K_\nu I_\nu'-K_\nu' I_\nu)$ . Each of the two products contributes $\tfrac{\nu}{2}\,\Gamma(\nu)/\Gamma(\nu+1)$ , and with $\Gamma(\nu+1)=\nu\Gamma(\nu)$ each equals $\tfrac12$ . Why: the gamma recursion is what makes the order-dependence cancel, so the answer cannot depend on $\nu$ .

Step 4. Add them: $xW\to \tfrac12+\tfrac12=1$ , so $C=1$ and $W[K_\nu,I_\nu]=1/x$ exactly. Why: this is the normalisation the factor $\pi/2$ in the definition of $K_\nu$ was engineered to produce, and it is what makes Green's-function amplitudes come out clean. *As a concrete check at $x=1$ : $I_0=1.26607$ , $K_0=0.42102$ , $I_0'=I_1=0.56516$ , $K_0'=-K_1=-0.60191$ ; then $K_0 I_0' - K_0' I_0 = 0.42102(0.56516)+0.60191(1.26607)=1.0000=1/x$ .*

Go deeper · Derivation C · 1 minThe screened Green's function in two dimensions

Derivation CThe screened Green's function in two dimensions

The goal is to derive the radial Green's function of the modified Helmholtz operator $-\nabla^2+\kappa^2$ in the plane and show it is $\tfrac{1}{2\pi}K_0(\kappa r)$ , exhibiting both why $K_0$ is selected over $I_0$ and how the delta source fixes the constant. This is the two-dimensional Yukawa (screened Coulomb) propagator, and the same steps carry to the Debye potential of a charged wire and to the correlation function of a massive free field.

Step 1. Away from the origin the radially symmetric equation $(-\nabla^2+\kappa^2)G=0$ reads $\tfrac1r\tfrac{\mathrm{d}}{\mathrm{d}r}(r\,G')-\kappa^2 G=0$ , which is the modified Bessel equation of order zero in the variable $\kappa r$ . Why: separation has already stripped the angular part, so the whole problem is order-zero modified Bessel, and its two solutions are $I_0(\kappa r)$ and $K_0(\kappa r)$ .

Step 2. Impose decay at infinity: $I_0$ grows like $e^{\kappa r}$ and is discarded, leaving $G(r)=A\,K_0(\kappa r)$ . Why: a screened field must vanish far from the source, which uniquely selects the Macdonald function — the opposite of an interior problem where $I_0$ is kept.

Step 3. Match the source using the small- $r$ behaviour $K_0(\kappa r)\sim -\ln(\kappa r/2)-\gamma$ , so $G\sim -A\ln r$ near the origin; integrating $-\nabla^2 G=\delta^{(2)}$ over a small disc turns the Laplacian into the outward flux $-\oint \partial_r G\,r\,\mathrm{d}\theta$ , which must equal $1$ . Why: the delta function is invisible in the homogeneous equation and only enters through this flux condition, which is how the constant $A$ is set.

Step 4. Evaluate the flux with $\mathrm{d}K_0/\mathrm{d}r\sim -1/r$ : $-2\pi r\,G'|_{r\to0}=2\pi A=1$ , hence $A=1/(2\pi)$ and $G(r)=\tfrac{1}{2\pi}K_0(\kappa r)$ . Why: this is the planar Yukawa propagator; its tail $e^{-\kappa r}$ is screening over the length $1/\kappa$ , its logarithmic core the point source. *As a consistency check on the limit: as $\kappa\to0$ , $K_0(\kappa r)\to-\ln(\kappa r)$ recovers the unscreened two-dimensional Coulomb Green's function $-\tfrac{1}{2\pi}\ln r$ up to an additive constant, confirming both the operator limit and the $1/2\pi$ normalisation.*

Limiting cases

The behaviour at the two ends of the real axis controls every application — regularity at the origin decides which function survives an interior boundary condition, and exponential decay at infinity decides which survives an exterior one.

Case	Result	Meaning
Small $x\to 0^+$ , $\nu>0$	$I_\nu\sim\dfrac{1}{\Gamma(\nu+1)}\left(\dfrac{x}{2}\right)^{\nu},\quad K_\nu\sim\dfrac{\Gamma(\nu)}{2}\left(\dfrac{2}{x}\right)^{\nu}$	$I_\nu$ vanishes like $x^\nu$ (regular); $K_\nu$ diverges like $x^{-\nu}$ (singular at the origin).
Small $x\to 0^+$ , $\nu=0$	$I_0(x)\to 1,\quad K_0(x)\sim-\ln(x/2)-\gamma$	The logarithmic singularity of $K_0$ mirrors the 2D Coulomb Green's function; $\gamma\approx0.5772$ is Euler–Mascheroni.
Large $x\to+\infty$	$I_\nu\sim\dfrac{e^{x}}{\sqrt{2\pi x}},\quad K_\nu\sim\sqrt{\dfrac{\pi}{2x}}\,e^{-x}$	Leading exponential rate is independent of $\nu$ ; the order enters only through subleading $(4\nu^2-1)/8x$ corrections.
Product at large $x$	$I_\nu(x)K_\nu(x)\sim\dfrac{1}{2x}$	The exponentials cancel, leaving the algebraic tail demanded by the exact Wronskian $W=1/x$ .
Half-integer $\nu=n+\tfrac12$	closed forms, e.g. $K_{1/2}(x)=\sqrt{\dfrac{\pi}{2x}}\,e^{-x}$	Elementary because the modified spherical Bessel functions terminate — the 3D Yukawa tail is exactly $e^{-x}$ .

Common trapAssuming the regular solution decays and the singular one grows, as with J and Y

The reflex carried over from the oscillatory case is backwards here. For $\{J_\nu,Y_\nu\}$ the regular $J_\nu$ is the one you usually keep. For the modified functions the roles split across the two ends of the axis: $I_\nu$ is regular at the origin but blows up at infinity, while $K_\nu$ is singular at the origin but decays at infinity. So an exterior screened-field problem keeps $K_\nu$ — the function badly behaved at the origin — because it is the only one that vanishes far away, the exact opposite of the $J_\nu$ instinct. Two further traps: the factor $\pi/2$ in $K_\nu$ is easy to drop, and at integer order $I_\nu$ and $I_{-\nu}$ become linearly dependent, so equation (4) is a $0/0$ limit taken by L'Hôpital — which is precisely where the $\ln(x/2)$ term in $K_n$ comes from.

Worked examples

1 · The Wronskian is exactly 1/x (confidence builder)

Using the tabulated values $I_0(1)=1.26607$ , $K_0(1)=0.42102$ , $I_1(1)=0.56516$ , $K_1(1)=0.60191$ , verify the identity $W[K_0,I_0](1)=1$ and explain why the answer is not a coincidence.

Step 1. Fix the derivatives from the recurrences. For order zero the ladder relations (6) collapse to $I_0'(x)=I_1(x)$ and $K_0'(x)=-K_1(x)$ — the zero-order functions have vanishing $\nu/x$ term, so their slopes are just the order-one functions (with a sign for $K$ ). At $x=1$ : $I_0'(1)=0.56516$ and $K_0'(1)=-0.60191$ .

Step 2. Assemble $W=K_0 I_0'-K_0' I_0 = K_0 I_1 + K_1 I_0$ . Substitute: $0.42102\times0.56516 = 0.23796$ and $0.60191\times1.26607 = 0.76207$ . Adding, $W(1)=\mathbf{1.00003}$ , which is $1/x$ at $x=1$ to five digits.

Step 3. Why it is exact: derivation B showed $xW=\tfrac12+\tfrac12=1$ from the gamma recursion, for every $\nu$ and every $x$ . The check is not luck but a probe of the normalisation — drop the $\pi/2$ in $K_\nu$ and this product would be $2/\pi$ instead of $1$ . The exact Wronskian $W[K_\nu,I_\nu]=1/x$ is what fixes the normalisation of the Macdonald function.

2 · A screened wire and the Debye length

A long charged wire sits in an electrolyte with Debye screening length $\lambda_D=1/\kappa$ . The radial potential is $\phi(r)\propto K_0(\kappa r)$ . By what factor does the potential fall between one and two Debye lengths from the wire, and how does that compare with an unscreened wire?

Step 1. Express distances in units of the Debye length, so $\kappa r=1$ at $r=\lambda_D$ and $\kappa r=2$ at $r=2\lambda_D$ . The potential ratio is then $\phi(2\lambda_D)/\phi(\lambda_D)=K_0(2)/K_0(1)$ .

Step 2. Insert the tabulated values $K_0(1)=0.42102$ and $K_0(2)=0.11389$ : the ratio is $0.11389/0.42102=\mathbf{0.2705}$ . The screened potential drops to about 27 % across a single extra Debye length — and because $K_0\sim\sqrt{\pi/2x}\,e^{-x}$ at larger argument, each further Debye length multiplies it by roughly $e^{-1}\approx0.37$ .

Step 3. Contrast the unscreened case, $\kappa\to0$ , where derivation C gives the bare logarithmic potential $\phi\propto-\ln r$ . From $r=\lambda_D$ to $2\lambda_D$ that changes only by $\ln 2\approx0.69$ — a slow logarithmic crawl, never a decay. Screening replaces the endless logarithmic tail of the 2D Coulomb field with an exponential cutoff on the scale $\lambda_D$ — exactly the physics $K_0$ encodes. The same $J_\nu$ -core, $K_\nu$ -cladding split governs a step-index optical fibre, with single-mode operation set by $V<2.405$ , the first zero of $J_0$ in the first figure.

Physics of the portrait

The modified Bessel functions carry two names between them. $I_\nu$ has no single owner — it is Bessel's function rotated — but the decaying second solution $K_\nu$ is universally the Macdonald function, after the Scottish mathematical physicist who singled it out while trying to understand how radio waves bend around the curve of the Earth.

Origins · 1 minHector Macdonald and how this was discovered

Hector Macdonald

1865–1935 · the Macdonald function K_ν and the diffraction of electric waves

Hector Munro Macdonald was born in Edinburgh, studied at Aberdeen and then Clare College, Cambridge, and was elected a Fellow of the Royal Society in 1901 for work on the mathematical theory of electric waves. In his 1899 paper "Zeroes of the Bessel functions" in the Proceedings of the London Mathematical Society he isolated and named the particular exponentially decaying second solution of the modified Bessel equation, giving its integral representation and behaviour, and $K_\nu(x)$ has borne his name ever since. His motivating problem was intensely practical: after Marconi's 1901 transatlantic transmission, physicists could not explain how signals reached beyond the geometric horizon, and the diffraction of waves around a conducting sphere — a problem whose radial part is exactly modified-Bessel in character — became a battleground he entered against Poincaré and Rayleigh. His diffraction formula was later superseded once the ionosphere was understood, but it was mathematically foundational. That a function now central to screened potentials, Yukawa propagators, and fibre optics entered physics through a question about long-distance radio is a fair emblem of how the modified Bessel functions travel.

History

1824

— Bessel's memoir on planetary perturbation establishes the functions J_n

1848

— steady heat flow in cylinders brings exponential radial solutions into use

1899

— Macdonald defines and names the decaying second solution K_ν

1922

— Watson's Treatise on the Theory of Bessel Functions fixes the modern I_ν, K_ν notation

1923

— Debye and Hückel's screening theory makes K_0 the kernel of the screened Coulomb potential

Going further

Texts. Arfken & Weber, Mathematical Methods for Physicists, 7th ed., Ch. 14 (Bessel functions, modified functions) · Stone & Goldbart, Mathematics for Physics, Ch. 9 (special functions and Green's functions) · Hassani, Mathematical Physics, 2nd ed., Ch. 14–15 · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 7 · Watson, A Treatise on the Theory of Bessel Functions, 2nd ed., for the definitive account. Key numerical anchors: series (2) reproduces $I_\nu=i^{-\nu}J_\nu(ix)$ via $(-1)^k i^{2k}=1$ ; the Wronskian satisfies $W\cdot x=1.0000$ at $I_0(1)=1.26607$ , $K_0(1)=0.42102$ , $I_1(1)=0.56516$ , $K_1(1)=0.60191$ ; the 2D Green's function constant $1/2\pi$ follows from flux integration and its $\kappa\to0$ log limit; the recurrence carries a difference sign (not a sum as for ordinary Bessel); and $K_0(2)/K_0(1)=0.2705$ , with $K_1(2)=0.13987$ .

Prerequisites. Bessel Functions of the First and Second Kind · The Gamma Function · Hankel Functions and Asymptotics · Recurrence Relations for Special Functions. Leads to. Spherical Bessel Functions · Spherical Harmonics · Generating Functions for Special Functions.

Ask the tutor: Ask me to walk through why the exterior of a screened wire keeps $K_0$ while the interior of a fibre keeps $J_0$ , and I will build the matching condition with you step by step.

Why are these functions called "modified" Bessel functions?

Because they solve Bessel's equation with the sign of the $x^2$ term reversed, which is exactly what the substitution $x\to ix$ does — it "modifies" Bessel's equation into one whose solutions are exponential rather than oscillatory. So $I_\nu$ and $K_\nu$ are literally Bessel functions of imaginary argument, $I_\nu(x)=i^{-\nu}J_\nu(ix)$ ; the name is shorthand for "Bessel functions on the imaginary axis". Everything structural about ordinary Bessel functions — series, recurrences, integral representations, a constant Wronskian — survives the rotation, which is why the two families are treated together.

When do I keep I_ν and when do I keep K_ν?

It is decided entirely by the boundary. Keep $I_\nu$ when the field must be finite at the origin (an interior problem — temperature on a cylinder axis, the field in a fibre core), because $I_\nu$ is the regular solution and $K_\nu$ diverges. Keep $K_\nu$ when the field must vanish far away (an exterior problem — a screened potential, an evanescent cladding), because $K_\nu$ decays and $I_\nu$ blows up. When both boundaries matter, as in an annulus, keep the combination $A\,I_\nu+B\,K_\nu$ and fix $A,B$ from the two conditions. The rule is opposite to the $J_\nu/Y_\nu$ habit, so state it explicitly each time.

Why does K_ν carry that strange π/2 and the combination I₋ν − I_ν?

Two independent jobs. The combination $I_{-\nu}-I_\nu$ is what produces a solution that decays at infinity: individually both $I_{\pm\nu}$ grow like $e^x$ , but their difference is arranged so the growing pieces cancel and only the decaying tail survives. The factor $\pi/2$ is a normalisation choice, and it is not arbitrary — it is precisely the constant that makes the Wronskian $W[K_\nu,I_\nu]$ come out to the clean value $1/x$ and the asymptotic form to $\sqrt{\pi/2x}\,e^{-x}$ . Change that prefactor and every Green's-function amplitude built from $K_\nu$ would carry an ugly compensating constant. The division by $\sin(\nu\pi)$ makes the definition fail at integer order, where a limit is needed.

What is the difference between K_ν and Y_ν — aren't both "second kind"?

Both are the second, irregular solution of their equations, but they behave oppositely at infinity. $Y_\nu$ oscillates and decays only algebraically, $\sqrt{2/\pi x}\,\sin(\dots)$ ; it is a standing wave. $K_\nu$ decays exponentially, $\sqrt{\pi/2x}\,e^{-x}$ ; it is a screened tail. They are related by continuation through the Hankel functions, $K_\nu(x)=\tfrac{\pi}{2}i^{\nu+1}H^{(1)}_\nu(ix)$ , so $K_\nu$ is the modified analogue of the outgoing Hankel function, not of $Y_\nu$ directly. "Second kind" refers to their role as the second independent solution, not to any similarity in behaviour.

Why does I_ν have no zeros while J_ν has infinitely many?

Look at the series. $J_\nu$ has an alternating sign $(-1)^k$ , so its terms fight and the function crosses zero infinitely often — those crossings are the Bessel zeros that quantise a drumhead's modes. $I_\nu$ is the same series with the sign removed, $(-1)^k\to+1$ , so every term is positive for $x>0$ and the sum can only climb; it never returns to zero. Structurally this is the difference between an oscillator ( $y''=-k^2 y$ ) and a runaway ( $y''=+\kappa^2 y$ ). It is why $I_\nu$ is the convenient interior profile in eigenvalue problems: finite and node-free right up to the boundary, where the quantisation condition is imposed.

How do the modified spherical Bessel functions fit in?

They are the half-integer-order cases dressed with a $\sqrt{\pi/2x}$ factor: $i_n(x)=\sqrt{\pi/2x}\,I_{n+1/2}(x)$ and $k_n(x)=\sqrt{\pi/2x}\,K_{n+1/2}(x)$ , solving the radial part of $(-\nabla^2+\kappa^2)$ in three dimensions rather than two. Because half-integer Bessel functions are elementary, these terminate into exponentials times polynomials — for instance $k_0(x)=(\pi/2)\,e^{-x}/x$ , precisely the Yukawa potential $e^{-\kappa r}/r$ . So the 3D screened point charge is a modified spherical Bessel function, just as the 2D screened line charge is $K_0$ .