Fourier Optics

Why Fourier optics matters

Fourier optics is what turns a lens from a piece of curved glass into a computer. It says that the act of focusing is literally the act of computing a Fourier transform — the operation at the heart of signal processing — at the speed of light and in two dimensions simultaneously. Once you accept that, a whole zoo of optical phenomena collapses into one idea: diffraction, image formation, resolution limits, spatial filtering, phase contrast, holography, matched-filter pattern recognition, and the entire theory of the microscope and telescope become special cases of linear systems theory applied to light.

The practical payoff is enormous. It tells the microscope designer that resolution is bought with numerical aperture, not magnification. It tells the lithographer at a semiconductor fab that shrinking a transistor means shrinking λ or raising NA, because Abbe's limit d = λ/(2·NA) is a wall. It gives astronomers deconvolution and adaptive optics, both grounded in the point spread function. And it gave engineers the 4f correlator — the first optical computers of the 1960s, which could recognize a shape in an aerial photograph by inserting a single glass filter into the Fourier plane. Every one of these follows from one sentence: the field in a lens's focal plane is the Fourier transform of the field entering it.

How it works, step by step

1. Light is a superposition of plane waves (the angular spectrum). Any field U(x, y) leaving an object plane can be decomposed into plane waves travelling in different directions. A plane wave with transverse wavevector (k_x, k_y) travels at angles sin θ_x = k_x/k, and it corresponds to a spatial frequency f_x = k_x/(2π) = sin θ_x / λ. Fine detail — sharp edges, closely spaced lines — means large angles, i.e. high spatial frequency. This decomposition is the angular spectrum, and it is a Fourier transform.

2. Free space propagates each plane wave with a phase. Over a distance z, a plane wave picks up a phase exp(ik_z·z) with k_z = √(k² − k_x² − k_y²). In the paraxial (small-angle) limit this becomes the Fresnel propagator, whose signature is a quadratic phase, exp(−iπλz(f_x² + f_y²)). Propagation smears the spectrum with this quadratic chirp; that chirp is exactly what a lens is built to cancel.

3. A thin lens multiplies by a quadratic phase. A converging lens of focal length f imposes a phase delay t(x, y) = exp(−ik(x² + y²)/2f) — thicker glass at the centre retards the wave more, which is what focuses it. Notice the sign: the lens's quadratic phase is the exact negative of free-space propagation's quadratic phase over one focal length.

4. The two quadratic phases cancel — leaving a pure Fourier transform. Place the object one focal length in front of the lens and observe one focal length behind it. Work through the Fresnel integral before the lens, multiply by the lens phase, then propagate the Fresnel integral after: the quadratic terms annihilate and the surviving integral is the plain Fourier transform. Concretely, the field in the back focal plane is

U_f(x', y') = (1 / iλf) · ∬ U_0(x, y) · exp[ −i(2π/λf)(x·x' + y·y') ] dx dy

which is the 2-D Fourier transform of the input U_0 evaluated at spatial frequencies f_x = x'/(λf) and f_y = y'/(λf). Position x' in the focal plane is spatial frequency, scaled by λf. This is why the back focal plane is called the Fourier plane.

5. A second lens inverse-transforms back to an image. A Fourier transform followed by another Fourier transform returns the original function (inverted). So a second lens after the Fourier plane rebuilds the image. That two-lens arrangement is the 4f system, and because the spectrum is physically laid out at the Fourier plane in the middle, you can reach in and edit it — this is spatial filtering.

6. The aperture truncates the spectrum → resolution and the PSF. No real lens is infinitely wide, so it can only collect spatial frequencies up to a cutoff set by its aperture. The pupil function P(x, y) — 1 inside the aperture, 0 outside — acts as a low-pass filter on the object spectrum. The image is therefore the inverse transform of the truncated spectrum, which is a blurred version of the object. Quantitatively, the coherent PSF is the Fourier transform of the pupil, and the incoherent PSF (what a microscope actually forms) is its squared magnitude, PSF = |F{P}|². For a circular pupil that is the Airy disc.

The 4f system and spatial filtering

The 4f processor is the canonical demonstration of Fourier optics and the ancestor of optical computing. Two lenses of focal length f are separated by 2f; the object sits 1f before the first lens, and the image forms 1f after the second lens — a total optical path of 4f, hence the name.

Plane	Distance from object	What lives there
Object (input)	0	The field U_0(x, y) to be processed
Lens 1	f	Forward Fourier transform
Fourier plane	2f	Spectrum F{U_0}; insert filter/mask here
Lens 2	3f	Inverse Fourier transform
Image (output)	4f	Filtered image (inverted)

Because the DC term (the average brightness, f = 0) sits at the exact centre of the Fourier plane and high frequencies sit at the edges, a physical mask in that plane multiplies the spectrum — and by the convolution theorem that multiplication is a convolution back in the image. The table below lists the classic filters.

Filter at Fourier plane	Effect on spectrum	Effect on image
Small central pinhole (low-pass)	Passes low f only	Blur / smooth — removes fine detail and noise
Central opaque stop (high-pass)	Blocks low f, passes high f	Edge enhancement; dark-field imaging
Slit along one axis	Passes one orientation	Keeps stripes in one direction only
λ/4 phase dot on the DC spot	Retards unscattered light by 90°	Zernike phase contrast — invisible cells become visible
Object's own conjugate spectrum	Multiplies by matched template	Correlation peak where the pattern occurs (VanderLugt correlator)

Frits Zernike won the 1953 Nobel Prize in Physics for the phase-contrast microscope, which is nothing more than a clever Fourier-plane filter that converts the invisible phase shifts imposed by a transparent cell into visible brightness variations. The VanderLugt matched filter (1964) let a beam of light search a photograph for a fingerprint or a tank in milliseconds — optical pattern recognition, decades before GPUs.

Resolution: the aperture is the whole story

Ernst Abbe, working for Carl Zeiss in Jena in 1873, realized why microscopes hit a wall. To resolve a periodic structure of period Λ, the lens must capture at least the zeroth and first diffraction orders, which leave the object at angle sin θ = λ/Λ. If the lens's half-angle of acceptance is θ_max, the finest resolvable period under matched illumination is

d = λ / (2 · NA),        NA = n · sin θ_max

where NA is the numerical aperture and n the refractive index of the medium (air ≈ 1.0, immersion oil ≈ 1.515). Equivalently, the highest spatial frequency the system transmits is f_cutoff = 2·NA/λ (incoherent) or NA/λ (coherent). Everything above that cutoff is physically absent from the image — no amount of magnification, contrast or post-processing can invent it, because those Fourier components never entered the lens.

Instrument / regime	λ	NA	Abbe limit d = λ/(2·NA)
Dry objective, green light	550 nm	0.95	≈ 290 nm
Oil-immersion objective	550 nm	1.40	≈ 200 nm
Blue light, oil immersion	450 nm	1.45	≈ 155 nm
193 nm ArF lithography (dry)	193 nm	0.93	≈ 104 nm
193 nm immersion lithography	193 nm	1.35	≈ 71 nm
13.5 nm EUV lithography	13.5 nm	0.33	≈ 20 nm

The closely related Rayleigh criterion for two point sources uses the Airy pattern: two points are "just resolved" when the first dark ring of one Airy disc lands on the peak of the other, giving an angular resolution θ = 1.22·λ/D for a telescope of diameter D, or a linear resolution 0.61·λ/NA at the object. This is why the 2.4 m Hubble mirror resolves ≈ 0.05 arcsec at 500 nm, and why the 39 m ELT will do vastly better — resolution scales as 1/D.

The point spread function and convolution imaging

An imaging system is, to excellent approximation, a linear, shift-invariant system. Feed it a single point of light and it returns not a point but a blurred blob — the point spread function h(x, y), the optical "impulse response." Because the system is linear, an arbitrary object is a sum of points, and each is imaged as a shifted, scaled copy of h. Adding them up is precisely the definition of convolution:

i(x, y) = ∬ o(ξ, η) · h(x − ξ, y − η) dξ dη  =  o(x, y) ⊗ h(x, y)

Here o is the ideal geometric image, h is the PSF, ⊗ denotes convolution, and i is the actual (blurred) image. Now apply the convolution theorem — convolution in space is multiplication in frequency:

I(f_x, f_y) = O(f_x, f_y) · H(f_x, f_y)

where uppercase letters are the Fourier transforms of the lowercase ones, and H = F{h} is the optical transfer function (OTF); its magnitude |H| is the modulation transfer function (MTF). The OTF is the Fourier transform of the PSF and, equivalently, the autocorrelation of the pupil. It is a low-pass filter: it passes low spatial frequencies with contrast near 1 and rolls off to exactly zero at the cutoff f_cutoff = 2·NA/λ. Blur, then, is nothing but attenuation of high spatial frequencies. This is the engine behind deconvolution (dividing out H to recover O) and behind every camera-lens MTF chart you have ever seen.

Symbols and units

• U(x, y) — complex scalar field amplitude (V/m or arbitrary units); intensity is |U|².
• λ — wavelength (m); in a medium, λ = λ_vacuum/n.
• f — lens focal length (m).
• k — wavenumber, k = 2π/λ (rad/m).
• f_x, f_y — spatial frequencies (cycles/m, often line pairs/mm), f_x = x'/(λf) = sin θ_x/λ.
• NA — numerical aperture, NA = n·sin θ_max (dimensionless).
• D — clear aperture diameter (m); d — smallest resolvable feature (m).
• P(x, y) — pupil function (dimensionless; 1 inside aperture, 0 outside).
• h, H — point spread function and optical transfer function (H = F{h}).

Worked example: focal length, aperture, and an Airy disc

Take a f = 200 mm lens with a clear aperture D = 25 mm, illuminated by a green HeNe-like λ = 543 nm laser. First, the spatial-frequency scale of the Fourier plane. A grating of period Λ = 10 µm diffracts to sin θ = λ/Λ = 0.543/10 = 0.0543, and the lens maps that angle to a spot at x' = f·tan θ ≈ f·sin θ = 200 mm × 0.0543 = 10.9 mm from the axis. Equivalently, f_x = 1/Λ = 100 cycles/mm lands at x' = λf·f_x = 543×10⁻⁶ mm × 200 mm × 100 mm⁻¹ = 10.9 mm. Halving the grating period doubles f_x and pushes the spot twice as far out — high frequency lives at the edge.

Now the PSF. The Airy disc radius (first dark ring) is

r_Airy = 1.22 · λ · f / D = 1.22 × 543×10⁻⁶ mm × 200 mm / 25 mm ≈ 5.3 µm

So a "perfect point" is smeared into a 5.3 µm-radius blob. The f-number is N = f/D = 8, and the Airy radius can also be written 1.22·λ·N = 1.22 × 0.543 µm × 8 ≈ 5.3 µm — the two forms agree. If we stop the lens down to D = 12.5 mm (N = 16), the Airy disc doubles to ≈ 10.6 µm: a smaller aperture chops off more high frequencies and blurs more. This is the quantitative face of "the aperture sets the resolution."

Common misconceptions

• "Magnification determines resolution." No — the numerical aperture does. Enlarging a blurred image only produces "empty magnification"; the missing high spatial frequencies were never collected by the lens and cannot be restored by scaling.
• "The Fourier plane only exists in special optical benches." Every imaging lens forms a Fourier transform at its back focal plane; the pupil (aperture stop) sits essentially there and is exactly what low-pass filters the image. It is not exotic — it is why every lens blurs.
• "A lens's Fourier transform requires laser light." The transform relation is a property of the geometry and holds for any field. Coherent illumination makes the amplitude transform directly observable; incoherent light instead makes the system linear in intensity, so imaging is convolution of intensities with |PSF|². Both are Fourier optics — the coherent case handles amplitudes and the OTF is the pupil; the incoherent case handles intensities and the OTF is the pupil autocorrelation.
• "The image is the transform of the object." Only the Fourier plane holds the transform. The image plane holds the inverse transform of the aperture-filtered spectrum — i.e. the object convolved with the PSF, which is nearly the object again, just band-limited.
• "Diffraction and imaging are different phenomena." They are the same Fourier mathematics. Fraunhofer diffraction of an aperture is its Fourier transform; a lens just brings that far-field pattern to a finite focal plane.
• "Spatial filtering adds information." It can only remove or re-weight spatial-frequency components that are already present. High-pass edge enhancement discards low frequencies; it never resurrects detail beyond the aperture cutoff.

A short history

The seeds are in Joseph Fourier's 1822 heat theory, but the optical connection begins with Ernst Abbe (1873), whose diffraction theory of the microscope explained resolution in terms of captured diffraction orders and gave Zeiss instruments their edge. Lord Rayleigh supplied the resolution criterion and much of the diffraction-integral machinery in the 1870s–90s. Frits Zernike (1934, Nobel 1953) turned Fourier-plane phase manipulation into phase-contrast microscopy. The systems-theory synthesis — treating optics as linear filters with transfer functions — came from work by Pierre-Michel Duffieux (1946) and was brought into engineering by Emil Wolf, Adolf Lohmann, Anthony VanderLugt and, definitively, Joseph W. Goodman, whose 1968 textbook Introduction to Fourier Optics made "a lens is a Fourier transformer" a canonical, teachable statement. From there it flows straight into holography (Gabor, 1948; Leith–Upatnieks, 1962), optical computing, and the resolution engineering behind every modern microscope, telescope and lithography stepper.