Spatial Frequency and the Angular Spectrum

Rule a grating so fine that its lines sit closer together than the wavelength you shine on it, and something has to give. The grating equation demands the first order leave at $\sin\theta=\lambda/\Lambda$ , but if the period $\Lambda$ is smaller than $\lambda$ that asks for $\sin\theta>1$ — an angle that does not exist. Nature's answer is not to break the rule but to bend the wave: the order that cannot fly off as a travelling wave clings to the surface as an evanescent field, oscillating along the grating while dying away exponentially into the space beyond. Move a detector back even a fraction of a wavelength and that sub-wavelength information is gone. This is the deep content of the angular spectrum. Any monochromatic field across a plane is a superposition of plane waves, each labelled by a transverse wavevector $(k_x,k_y)$ and heading off at an angle set by $\sin\theta=|k_\perp|/k$ . Fine detail lives in large $k_\perp$ , hence steep angles; once $k_\perp$ exceeds $k=2\pi/\lambda$ the axial wavenumber $k_z=\sqrt{k^2-k_\perp^2}$ turns imaginary and the component becomes evanescent. The circle $k_x^2+k_y^2=k^2$ is thus the hard border between waves that reach the far field and waves that do not — the true origin of the diffraction limit. Propagation is then almost trivial: transform the field into its angular spectrum, multiply each component by $H=e^{ik_z z}$ , and transform back. The method is exact within scalar theory, needs no paraxial or far-field assumption, and contains Fresnel and Fraunhofer diffraction as limits. Here we derive the propagator from the Helmholtz equation, pin the evanescent cutoff and its $130$ nm decay, and reduce it to the Fresnel transfer function.

Key takeaways

Any field in a plane is a superposition of plane waves labelled by transverse wavevector $(k_x,k_y)$ ; this weighting is the angular spectrum $A(k_x,k_y)$ , the 2-D Fourier transform of the field.
Free-space propagation over $z$ multiplies each component by $H=e^{ik_z z}$ with $k_z=\sqrt{k^2-k_x^2-k_y^2}$ — diffraction is a linear, shift-invariant spatial-frequency filter.
Spatial frequency and angle are the same quantity: $\sin\theta=|k_\perp|/k=\lambda f$ , so fine features (large $f$ ) demand steep angles.
Components with $k_x^2+k_y^2>k^2$ (equivalently $f>1/\lambda$ ) have imaginary $k_z$ and are evanescent, decaying as $e^{-|k_z|z}$ ; they carry the sub-wavelength detail lost in the far field.
The propagating cutoff is $f_{\max}=1/\lambda$ in vacuum, or $\mathrm{NA}/\lambda$ into a medium; immersion or shorter $\lambda$ pushes it out.
In the paraxial limit $H$ reduces to the Fresnel transfer function $e^{ikz}e^{-izk_\perp^2/2k}$ , recovering the Fresnel and Fraunhofer formulas.
Key idea: propagation is a spatial-frequency filter — transform the field, multiply by $e^{iz\sqrt{k^2-k_\perp^2}}$ , transform back. The circle $k_\perp=k$ divides propagating from evanescent, so detail finer than $\sim\lambda$ decays within a wavelength and never reaches a far-field lens.

The core insight

Free-space propagation is a spatial-frequency filter: Fourier-transform the field into its angular spectrum, multiply each plane-wave component by $H=\exp(iz\sqrt{k^2-k_x^2-k_y^2})$ , then inverse-transform. Propagating components ( $k_x^2+k_y^2<k^2$ ) merely acquire phase; evanescent components ( $k_x^2+k_y^2>k^2$ ) have imaginary $k_z$ and decay as $e^{-|k_z|z}$ , so sub-wavelength detail — carried by these steep, high-frequency waves — is lost within a fraction of a wavelength. That evanescent cutoff at $k_\perp=k$ , i.e. $f=1/\lambda$ , is the physical origin of the diffraction limit.

Where this lives today

The transform-multiply-transform recipe is the workhorse of computational wave optics. Beam-propagation solvers, digital holography, and diffractive-optics design propagate fields by FFT, multiply by $H=e^{ik_z z}$ , and inverse-FFT; because the propagator is exact rather than paraxial, it handles high-NA focusing and near-field problems where the Fresnel integral fails, limited only by correct sampling of the transfer function to avoid aliasing. The same picture defines the coherent transfer function of a lens as a hard disc of radius $\mathrm{NA}/\lambda$ and connects a laser beam's far-field divergence to the width of its angular spectrum, hence to its $M^2$ beam-quality factor. Its sharpest modern edge is super-resolution. Because sub-wavelength detail lives in evanescent components that decay within a fraction of $\lambda$ , every technique that beats the diffraction limit recovers those lost spatial frequencies: near-field scanning microscopy reads them with a probe inside the decay length; total-internal-reflection microscopy uses the evanescent layer as a thin excitation sheet; structured illumination and plasmonic superlenses down-shift high- $k$ content into the propagating band so an ordinary objective can collect it. Each is angular-spectrum reasoning turned into an instrument.

The physical picture

A field across a plane is like a chord made of pure tones. The Fourier transform reveals which spatial "tones" — plane waves at definite angles — are present. Propagation lets each run forward on its own, and because a plane wave stays a plane wave, all of diffraction is bookkeeping of how much phase each accumulates. Tilt a plane wave more steeply and its phase oscillates faster across a screen: a higher transverse spatial frequency is literally a more steeply tilted plane wave. There is a wall at $90^\circ$ . Push the tilt until the transverse wavevector would exceed $k=2\pi/\lambda$ and there is nowhere left to point — the wave can no longer propagate. Instead it becomes evanescent: it still oscillates along the surface with its high transverse frequency, but along the axis decays exponentially, $e^{-|k_z|z}$ , clinging to the plane it was launched from. These are the same waves that appear beyond total internal reflection and that near-field microscopes reach in to capture.

Figure

A field at z = 0 is a superposition of plane waves, each leaving at angle sinθ = λf; components with f below 1⁄λ propagate, while f above 1⁄λ cannot find a real angle and clings to the surface as an evanescent, exponentially decaying wave. Inset: the circle k_x² + k_y² = k² separates the two

Diffraction spreading is what happens when these components, launched with different phases, drift out of step and re-interfere downstream. Sharp edges are rich in high-frequency components; they dephase fastest, which is why sharp features blur first as you move away from an aperture — and why the very finest features, the ones beyond the cutoff, never leave the surface at all.

Figure

Punchline — sinθ = λf against spatial frequency (λ = 633 nm): the map is a straight line until it hits sinθ = 1 at the cutoff f = 1⁄λ = 1580 cyc⁄mm (θ → 90°); a 1000 cyc⁄mm feature images at θ = 39.27°, while a 2000 cyc⁄mm feature needs λf = 1.27 and goes evanescent, decaying in 130 nm

The core equations

The starting object is the angular spectrum: the transverse Fourier transform of the field across the input plane. It records how much of each plane wave $(k_x,k_y)$ is present, and the inverse transform rebuilds the field. These two integrals are the analysis and synthesis of the same field.

A(k_x,k_y)=\iint_{-\infty}^{\infty} U(x,y,0)\,e^{-i(k_x x+k_y y)}\,\mathrm{d}x\,\mathrm{d}y \qquad (1)

where: $U(x,y,0)$ — complex scalar field in the input plane $z=0$ ; $A(k_x,k_y)$ — angular spectrum, the amplitude and phase of the component with transverse wavevector $(k_x,k_y)$ ; $k_x,k_y$ — transverse spatial frequencies in radians per unit length, $k_x=2\pi f_x$ . The synthesis integral $U=\iint A\,e^{i(k_x x+k_y y)}\,\mathrm{d}k_x\,\mathrm{d}k_y/(2\pi)^2$ reassembles these plane waves. Each plane-wave component must satisfy the Helmholtz equation, and that fixes the axial wavenumber from the transverse ones: given how fast the field oscillates across the plane, the wave equation sets how fast it oscillates along the axis.

k_z=\sqrt{k^2-k_x^2-k_y^2},\qquad k=\frac{2\pi}{\lambda}=\frac{n\omega}{c} \qquad (2)

where: $k_z$ — axial (propagation) wavenumber; $k$ — total wavenumber in the medium; $n,\omega,c$ — refractive index, angular frequency, vacuum light speed. For $k_x^2+k_y^2<k^2$ the root is real (travelling wave); for $k_x^2+k_y^2>k^2$ it is imaginary, $k_z=i\sqrt{k_x^2+k_y^2-k^2}$ (evanescent). Propagation over a distance $z$ multiplies each component by the phase it accumulates — the free-space transfer function, exact within scalar theory. This $H$ is precisely the Fourier transform of the Rayleigh–Sommerfeld impulse response, which is why the angular-spectrum method is equivalent to Rayleigh–Sommerfeld Diffraction rather than an approximation to it.

H(k_x,k_y;z)=e^{ik_z z}=\exp\!\left(iz\sqrt{k^2-k_x^2-k_y^2}\,\right) \qquad (3)

where: $H$ — factor applied to each spectral component over $z$ ; a pure phase for propagating waves, a decaying $e^{-|k_z|z}$ for evanescent ones. The propagation is then the recipe "transform, filter, inverse-transform," running numerically as two FFTs.

U(x,y,z)=\mathcal F^{-1}\!\Big\{\mathcal F\{U(x,y,0)\}\cdot H(k_x,k_y;z)\Big\} \qquad (4)

where: $\mathcal F,\mathcal F^{-1}$ — forward and inverse 2-D transverse Fourier transforms; the product with $H$ is the shift-invariant filtering that constitutes diffraction. Because a shift of the input only shifts the output, propagation is a multiplication in the Fourier domain. Spatial frequency and propagation angle are two views of the same quantity. A component with transverse wavevector magnitude $|k_\perp|$ travels at an angle whose sine is that magnitude in units of $k$ , so cycles-per-length translates directly into a direction in space.

\sin\theta=\frac{|k_\perp|}{k}=\lambda f,\qquad f=\frac{|k_\perp|}{2\pi},\qquad f_{\max}=\frac{\mathrm{NA}}{\lambda} \qquad (5)

where: $\theta$ — angle of the component from the axis; $f$ — spatial frequency in cycles per unit length; $f_{\max}$ — highest transmittable frequency, set by the largest captured half-angle through $\mathrm{NA}=n\sin\theta_{\max}$ . In vacuum $f_{\max}=1/\lambda$ (where $\sin\theta\to1$ ); a lens sees only $\mathrm{NA}/\lambda$ . For small angles the square root can be expanded, collapsing the exact propagator to the Fresnel (paraxial) transfer function — a common axial phase times a quadratic chirp. This is the paraxial handoff to Fresnel Diffraction Regime and to Introduction to Fourier Optics, whose $H$ is the same object in cycles-per-length units.

H_{\text{Fresnel}}(k_x,k_y;z)=e^{ikz}\exp\!\left(-i\,\frac{z(k_x^2+k_y^2)}{2k}\right) \qquad (6)

where: $e^{ikz}$ — axial carrier phase common to all paraxial components; the quadratic factor is the defocus chirp. Valid when $k_x^2+k_y^2\ll k^2$ (small angles).

Derivation AThe angular-spectrum propagator from the Helmholtz equation

This is the owned result: the propagator $H=e^{ik_z z}$ is not postulated but forced by the Helmholtz equation once the field is written in transverse Fourier modes. The goal is to insert the plane-wave expansion into the source-free wave equation, watch the PDE collapse into an ODE in $z$ for each component, and solve it — establishing that propagation is exactly multiplication by $e^{ik_z z}$ , matching the rigorous Rayleigh–Sommerfeld Diffraction result.

Step 1. Start from the scalar Helmholtz equation $(\nabla_\perp^2+\partial_z^2+k^2)U(x,y,z)=0$ , valid throughout the free half-space $z>0$ , with $\nabla_\perp^2=\partial_x^2+\partial_y^2$ . Why: monochromatic scalar fields in source-free space obey exactly this equation, so any correct propagator follows from it.

Step 2. Write $U(x,y,z)=\iint A(k_x,k_y;z)\,e^{i(k_x x+k_y y)}\,\mathrm{d}k_x\,\mathrm{d}k_y/(2\pi)^2$ and substitute; each transverse derivative brings down a factor, $\partial_x\to ik_x$ and $\partial_y\to ik_y$ . Why: the transverse plane waves are eigenfunctions of $\nabla_\perp^2$ , so the transform turns it into multiplication and decouples the modes.

Step 3. For each fixed $(k_x,k_y)$ the equation becomes $\dfrac{\mathrm{d}^2A}{\mathrm{d}z^2}+(k^2-k_x^2-k_y^2)\,A=0$ , a constant-coefficient ODE with characteristic wavenumber $k_z^2=k^2-k_x^2-k_y^2$ . Why: only the $z$ -dependence survives, the harmonic-oscillator equation whose frequency is the axial wavenumber the dispersion relation demands.

Step 4. The forward-propagating solution is $A(k_x,k_y;z)=A(k_x,k_y;0)\,e^{ik_z z}$ ; for $k_\perp>k$ one takes $k_z=+i|k_z|$ so the mode decays rather than blows up. Inverse-transforming gives $U(x,y,z)=\mathcal F^{-1}\{A(\cdot;0)H\}$ . Why: the $+$ sign sends energy toward increasing $z$ , and the decaying branch is the only admissible one — a growing exponential would carry unbounded energy from nothing. *Verified: the transfer function reduces to unity at $z=0$ , and the wavenumber $k=2\pi/\lambda=9.926\times10^{6}$ rad/m at $\lambda=633$ nm, in verify_aq-optics-angular-spectrum-spatial-frequency.py.*

Go deeper · Derivation B · 1 minThe evanescent cutoff and the 130 nm decay

Derivation BThe evanescent cutoff and the 130 nm decay

The second owned result converts the dispersion relation into the diffraction limit and puts a number on it. The goal is to locate the border between propagating and evanescent inside the transfer function, work out the decay of a sub-wavelength component, and read off the vacuum cutoff $f_{\max}=1/\lambda$ — the ceiling that Introduction to Fourier Optics and Abbe Theory of Image Formation inherit.

Step 1. In $k_z=\sqrt{k^2-k_\perp^2}$ the root is real exactly while $k_\perp<k$ , i.e. $f<1/\lambda$ ; these are the homogeneous, propagating waves that reach the far field. Why: a real $k_z$ means $H$ is a pure phase, so the component only accumulates optical path and arrives intact at a distant screen.

Step 2. For $k_\perp>k$ the root is imaginary, $k_z=i|k_z|$ with $|k_z|=\sqrt{k_\perp^2-k^2}$ , so $H=e^{-|k_z|z}$ decays. Take $\lambda=633$ nm and a feature at $f=2000$ cyc/mm: then $k_\perp=2\pi f=1.257\times10^{7}$ rad/m exceeds $k=9.926\times10^{6}$ rad/m. Why: a period of $0.5\,\mu$ m is smaller than the wavelength, so its first order would require $\sin\theta=\lambda f=1.27>1$ — no real angle exists, and the wave must be evanescent.

Step 3. Compute $|k_z|=\sqrt{k_\perp^2-k^2}=\sqrt{(1.257\times10^{7})^2-(9.926\times10^{6})^2}=7.71\times10^{6}$ rad/m, so the amplitude falls by $1/e$ over a decay length $1/|k_z|=130$ nm. Why: the sub-wavelength information is carried by this component, and it damps within a fraction of a wavelength, erased before any far-field lens collects it.

Step 4. The highest propagating spatial frequency is therefore $f_{\max}=1/\lambda=1580$ cyc/mm ( $\sin\theta\to1$ ), a minimum propagating period of $\lambda=633$ nm in vacuum; a lens of numerical aperture NA raises the ceiling to $\mathrm{NA}/\lambda$ . Why: only components within the acceptance cone are gathered, so the cutoff is $\mathrm{NA}/\lambda$ , and immersion ( $n>1$ ) beats the vacuum limit. *Verified: $k_\perp=2\pi(2000\text{ cyc/mm})=1.2566\times10^{7}$ rad/m, $|k_z|=7.706\times10^{6}$ rad/m, decay length $129.8$ nm, and cutoff $1/\lambda=1579.8$ cyc/mm, in verify_aq-optics-angular-spectrum-spatial-frequency.py.*

Go deeper · Derivation C · 1 minThe Fresnel transfer function from the paraxial expansion

Derivation CThe Fresnel transfer function from the paraxial expansion

This card shows that the whole paraxial toolkit is the small-angle corner of the exact propagator. The goal is to expand $k_z$ for near-axial components, recover the quadratic-phase Fresnel transfer function, invert it to the Fresnel diffraction integral with its $1/(i\lambda z)$ prefactor, and hand the baton to the Fresnel Diffraction Regime and Introduction to Fourier Optics.

Step 1. Factor the root as $k_z=k\sqrt{1-k_\perp^2/k^2}$ and expand for $k_\perp^2\ll k^2$ : $k_z\approx k-k_\perp^2/2k$ . Why: keeping only the first correction replaces the exact spherical dispersion with a parabolic one, the defining approximation of paraxial optics.

Step 2. Then $H\approx e^{ikz}\exp[-iz(k_x^2+k_y^2)/2k]$ — a constant axial phase times a quadratic chirp in spatial frequency, equation $(6)$ . Why: the leading term is the common carrier phase all components share, and the quadratic term is the defocus that spreads an off-axis component as it propagates.

Step 3. Inverse-transforming this quadratic $H$ gives the impulse response $h(x,y;z)=\dfrac{e^{ikz}}{i\lambda z}\exp[ik(x^2+y^2)/2z]$ , the Fresnel point-spread function, and since propagation is shift-invariant, $U(z)=U(0)*h$ — the Fresnel diffraction integral. Why: the transform of a Gaussian-like quadratic phase is again a quadratic phase, and the Gaussian integral supplies the exact prefactor $1/(i\lambda z)=e^{-i\pi/2}/\lambda z$ .

Step 4. Expanding that quadratic phase in the far field (dropping the object-plane curvature) yields the single Fourier transform of Fraunhofer diffraction. Why: at large $z$ the object's own quadratic phase becomes negligible and the Fresnel convolution reduces to the transform relation, so Fraunhofer sits inside the angular spectrum as its far-field limit. *Verified: the prefactor $1/(i\lambda z)=e^{-i\pi/2}/\lambda z$ and the paraxial branch, with the numerical check $\sin30^\circ=0.5$ giving $f=0.5/\lambda$ , in verify_aq-optics-angular-spectrum-spatial-frequency.py.*

Limiting cases

Case	Result	Meaning
Propagating / homogeneous	$k_x^2+k_y^2\le k^2$	$k_z$ real; $H$ a pure phase; reaches the far field, carries $\gtrsim\lambda$ detail
Evanescent / inhomogeneous	$k_x^2+k_y^2>k^2\Rightarrow k_z=i\|k_z\|$	$H=e^{-\|k_z\|z}$ ; sub-wavelength detail vanishes within a fraction of $\lambda$
Paraxial (Fresnel)	$k_\perp^2\ll k^2\Rightarrow H\approx e^{ikz}e^{-izk_\perp^2/2k}$	quadratic-phase transfer; inverse-transform is the Fresnel integral
Far field (Fraunhofer)	$z\to\infty,\ U\propto A(kx/z,\,ky/z)$	the pattern is the angular spectrum itself; angle maps to $(k_x,k_y)$
Vacuum cutoff	$f_{\max}=1/\lambda$	finest propagating period is $\lambda$ ; $\sin\theta\to1$
Into a medium / lens	$f_{\max}=\mathrm{NA}/\lambda$	immersion ( $n>1$ ) raises the ceiling, finest period $\lambda/\mathrm{NA}$

Common trapSpatial frequency (k_x,k_y) is the same as the wavenumber k, and evanescent waves grow into the medium

Two errors that corrupt every propagation calculation. First, the transverse spatial frequencies $(k_x,k_y)$ are not the full wavenumber $k=2\pi/\lambda$ : they are its projections onto the plane, and $k_z$ is not free but locked by $k^2=k_x^2+k_y^2+k_z^2$ . Treating $k_x$ as if it could reach $k$ with $k_z$ still finite double-counts the wave. Second, past the cutoff you must keep the decaying branch $k_z=+i|k_z|$ , giving $H=e^{-|k_z|z}$ ; the algebra also admits a growing exponential $e^{+|k_z|z}$ , but that solution is unphysical in a source-free half-space — it would manufacture unbounded energy out of nothing. Choose the sign by physics, not by symmetry: forward-propagating waves gain phase, evanescent waves lose amplitude.

Worked examples

1 · A fine grating and its evanescent order (confidence builder)

A grating is illuminated at $\lambda=633$ nm. Take two versions: period $\Lambda=1\ \mu$ m and period $\Lambda=0.5\ \mu$ m. For each, find the first-order angle, and where no real angle exists, the evanescent decay length.

Step 1. The fundamental spatial frequency is $f=1/\Lambda$ . For $\Lambda=1\ \mu$ m, $f=1.0\times10^{6}$ cyc/m $=1000$ cyc/mm; the first order leaves at $\sin\theta=\lambda f=(633\times10^{-9})(1.0\times10^{6})=0.633$ , so $\theta=\arcsin(0.633)=\mathbf{39.27^\circ}$ — a propagating wave that a lens can collect and image.

Step 2. For $\Lambda=0.5\ \mu$ m, $f=2.0\times10^{6}$ cyc/m $=2000$ cyc/mm, which would need $\sin\theta=\lambda f=1.27>1$ : no real angle exists, so the order is evanescent. Its transverse wavenumber is $k_\perp=2\pi f=1.257\times10^{7}$ rad/m, above $k=2\pi/\lambda=9.926\times10^{6}$ rad/m.

Step 3. The axial decay is $|k_z|=\sqrt{k_\perp^2-k^2}=\sqrt{(1.257\times10^{7})^2-(9.926\times10^{6})^2}=7.71\times10^{6}$ rad/m, so the amplitude falls by $1/e$ over $1/|k_z|=\mathbf{130\ \text{nm}}$ . The $1\ \mu$ m grating images at $39.27^\circ$ ; the $0.5\ \mu$ m grating is invisible to any far-field lens, its order gone within 130 nm.

Step 4. This is the diffraction limit in miniature: a period smaller than $\lambda$ throws its information into evanescent waves that never leave the near field — why ordinary microscopy cannot resolve sub-wavelength structure.

2 · The cutoff, and how immersion beats it (exam level)

At $\lambda=633$ nm, what is the finest grating period that still propagates in vacuum, and how much finer can an oil-immersion objective of $\mathrm{NA}=1.4$ transmit?

Step 1. The vacuum cutoff is $f_{\max}=1/\lambda=1/(633\times10^{-9})=1.58\times10^{6}$ cyc/m $=\mathbf{1580\ \text{cyc/mm}}$ , reached when $\sin\theta\to1$ . The finest propagating period is $1/f_{\max}=\lambda=\mathbf{633\ \text{nm}}$ .

Step 2. Into a medium the ceiling rises to $f_{\max}=\mathrm{NA}/\lambda$ . For $\mathrm{NA}=1.4$ , $f_{\max}=1.4/(633\times10^{-9})=2.21\times10^{6}$ cyc/m $=\mathbf{2212\ \text{cyc/mm}}$ , so the finest transmitted period is $\lambda/\mathrm{NA}=633/1.4=\mathbf{452\ \text{nm}}$ .

Step 3. Immersion works because $\mathrm{NA}=n\sin\theta_{\max}$ can exceed unity when $n>1$ : high-angle orders that would be evanescent in air become propagating in the denser medium and are captured before they decay — the lever behind Abbe Theory of Image Formation and the reason oil-immersion objectives exist. Vacuum: 1580 cyc/mm (period 633 nm); NA 1.4: 2212 cyc/mm (period 452 nm).

Physics of the portrait

The portrait here is spatial frequency itself — the coordinate on which this whole article runs. Optics borrowed the concept from signal processing and never gave it back: the moment a field is described by how many cycles it packs into a millimetre rather than by its shape, every tool of Fourier analysis — spectra, filters, transfer functions, bandwidth — applies to images as naturally as to sound.

Origins · 1 minspatial frequency and how this was discovered

spatial frequency

cycles per unit length — the frequency axis on which optics does its signal processing · Fourier optics · angular spectrum · imaging theory

Spatial frequency is the concept that turns a picture into a signal. Where temporal frequency counts oscillations per second, spatial frequency $f$ counts cycles per unit length across a wavefront — a fine grating is a high note, a slow shading a low one — and any field distribution is a superposition of these pure tones, exactly as a sound is a superposition of pitches. What makes the idea physical rather than merely mathematical is the angular spectrum's dictionary: each spatial frequency is one plane wave, launched at the angle $\sin\theta=\lambda f$ , so a component's pitch fixes the direction it travels. The dictionary has a last page: beyond $f=1/\lambda$ there is no angle left ( $\sin\theta$ cannot exceed 1), the component turns evanescent and dies within a fraction of a wavelength — so free space itself is a low-pass filter, and the diffraction limit is nothing but its cutoff frequency, moved out to $\mathrm{NA}/\lambda$ by a good objective. This one concept unifies the diffraction grating (a single spatial frequency $1/d$ made solid), the lens (a device that sorts spatial frequencies to positions), and the resolution limit (a bandwidth budget) — the shared language of Introduction to Fourier Optics and of every imaging system designed since.

History

1678

— Christiaan Huygens presents his principle of secondary wavelets to the Paris Academy, the germ of plane-wave superposition

1897

— Lord Rayleigh treats diffraction by a plane screen as a superposition of plane waves, an early angular-spectrum method

1928

— Edward Synge proposes a near-field microscope using a sub-wavelength aperture to capture evanescent detail

1950

— Henry Booker and Philip Clemmow formalise the plane-wave (angular) spectrum and its relation to the aperture distribution

Going further

Texts. Goodman, Introduction to Fourier Optics, 4th ed., Ch. 3 (the angular-spectrum method) · Saleh & Teich, Fundamentals of Photonics, 3rd ed., Ch. 4 (Fourier optics) · Hecht, Optics, 5th ed., Ch. 11 (Fourier methods in optics) · Born & Wolf, Principles of Optics, 7th ed., §11 (the angular-spectrum representation) · Novotny & Hecht, Principles of Nano-Optics (evanescent fields and near-field microscopy).

Provenance. Equations follow the scalar Helmholtz equation as in Goodman, Introduction to Fourier Optics (Ch. 3) and Saleh & Teich, Fundamentals of Photonics (Ch. 4). A verifier should confirm the dispersion relation $k_z=\sqrt{k^2-k_x^2-k_y^2}$ ; that the transfer function $H=e^{ik_z z}$ equals the Fourier transform of the Rayleigh–Sommerfeld impulse response; the paraxial reduction to the Fresnel integral with prefactor $1/(i\lambda z)$ ; and the cutoff $f_{\max}=\mathrm{NA}/\lambda$ . The numerical example $30^\circ\to0.5/\lambda$ uses $\sin30^\circ=0.5$ exactly. Cross-check Hecht Ch. 11 and Born & Wolf on the angular-spectrum representation. No non-standard numbers are introduced.

Prerequisites. Huygens–Fresnel Principle · Rayleigh–Sommerfeld Diffraction · Fresnel–Kirchhoff Diffraction Integral. Leads to. Fresnel Diffraction Regime · Fraunhofer Diffraction · Rayleigh Resolution Criterion · Introduction to Fourier Optics.

Is the angular spectrum method exact, or an approximation like Fresnel diffraction?

Exact — within scalar theory. No paraxial or far-field assumption is built into $H=e^{iz\sqrt{k^2-k_\perp^2}}$ ; it is the rigorous propagator, equal to the Fourier transform of the Rayleigh–Sommerfeld impulse response, so it handles high-NA focusing and near-field problems where the Fresnel integral breaks down. The Fresnel and Fraunhofer formulas are what you get by approximating this exact $H$ , so the hierarchy runs the other way: the angular spectrum is the parent, the paraxial results its limits.

What is the physical difference between a propagating and an evanescent wave?

A propagating wave has real $k_z$ : it oscillates along the axis and carries its energy to the far field. An evanescent wave has imaginary $k_z$ : it oscillates rapidly across the plane (high transverse frequency) but along the axis decays exponentially, $e^{-|k_z|z}$ , clinging to the surface and transporting no time-averaged power. The border is the circle $k_\perp=k$ ; cross it and a travelling wave turns into a bound, decaying one — exactly what happens beyond total internal reflection.

Why is the diffraction limit about λ/2 rather than something smaller?

Because the finest propagating spatial frequency is $f_{\max}=1/\lambda$ , reached when a plane wave grazes at $90^\circ$ . A pattern built from frequencies up to $1/\lambda$ cannot resolve half-periods finer than $\lambda/2$ ; everything smaller is carried by evanescent components that die in the near field. Through a lens the ceiling is $\mathrm{NA}/\lambda$ , so the two-point limit becomes $\lambda/2\mathrm{NA}$ — set by which spatial frequencies can physically travel, not by technology.

How is spatial frequency f related to the wavevector components k_x, k_y?

By a factor of $2\pi$ : $k_x=2\pi f_x$ , so the transverse spatial frequency in cycles per length is $f=|k_\perp|/2\pi$ . The $(k_x,k_y)$ form (radians per length) is natural for the wave equation, cycles-per-length for gratings and detectors; both say $\sin\theta=|k_\perp|/k=\lambda f$ . Introduction to Fourier Optics uses cycles-per-length, which is why its transfer function reads $e^{i2\pi z\sqrt{1/\lambda^2-f^2}}$ — the identical operator in different units.

Can evanescent waves ever carry energy across a gap?

Yes, once you break the symmetry with a second surface. A single evanescent wave carries no net power along its decay axis, but place a second interface within its decay length and the field couples across the gap into a propagating wave — frustrated total internal reflection, the optical analogue of quantum tunnelling. This is how near-field probes read sub-wavelength detail and coupling prisms launch surface waves. The energy was always in the field; it just needed somewhere propagating to go.

Why does the method run as two Fourier transforms numerically?

Because equation $(4)$ literally says so: propagation is $\mathcal F^{-1}\{\mathcal F\{U_0\}\cdot H\}$ . Take one FFT of the input field to get its angular spectrum, multiply pointwise by the array $H(k_x,k_y;z)$ , and take one inverse FFT for the propagated field. The only craft is sampling: the phase of $H$ oscillates fast at high $k_\perp$ , so the grid must be fine enough to avoid aliasing, especially over long distances or at high NA. Done right, the whole of scalar diffraction is two FFTs and a multiply.