Optical Transfer Function

Point a camera at a page of ever-finer stripes and watch what happens: the coarse bars come through crisp and black-on-white, the medium bars turn grey, and past some fineness the stripes dissolve into a flat tone. The lens has not simply "blurred" the page — it has applied a well-defined filter, passing low spatial frequencies at nearly full strength and squeezing the contrast out of high ones until, at a sharp cutoff, it transmits nothing. The optical transfer function (OTF) is that filter written down: the complex number by which the imaging system multiplies each spatial-frequency component of the ideal image. The framework is linear-systems theory borrowed wholesale from electronics. If a system is linear and shift-invariant across the field — the same blur wherever a point lands — then the image is the ideal image convolved with the system's impulse response, the point-spread function. Convolution in space is multiplication in frequency, so the whole imaging behaviour collapses into one function of spatial frequency: the Fourier transform of the point-spread function, normalized to unity at zero frequency. Its modulus is the modulation transfer function (MTF) — how much fringe contrast survives — and its phase is the phase transfer function (PTF) — how features shift. For incoherent light (the everyday case: fluorescence, daylight, an LED-lit page) the OTF turns out to be the autocorrelation of the pupil: slide a copy of the aperture across itself, and the fractional overlap area at each shift is the MTF. That single geometric picture yields the diffraction-limited curve, its cutoff at twice the coherent limit, and the connection to Abbe's resolution — all without evaluating a diffraction integral. This article builds the OTF from the point-spread function, derives the pupil-overlap formula and the circular-aperture MTF, and shows where the numbers land in a real lens.

Key takeaways

The OTF is the complex spatial-frequency response of an imaging system: it multiplies the Fourier spectrum of the ideal image to give the spectrum of the actual image.
OTF = normalized Fourier transform of the point-spread function; its modulus is the MTF (contrast survival), its argument the PTF (feature shift / contrast reversal).
For incoherent illumination the OTF is the normalized autocorrelation of the pupil — the fractional overlap area of two shifted copies of the aperture.
The diffraction-limited incoherent cutoff is `f_c = 2·NA/λ`, exactly twice the coherent cutoff `NA/λ`, because autocorrelation doubles the pupil's frequency support.
A negative OTF is not an error: it means contrast reversal (a π phase), the "spurious resolution" seen beyond a defocus zero.
Key idea: the incoherent MTF at spatial frequency f is the fractional overlap of the pupil with a copy of itself shifted by λ·d_i·f — turning a diffraction calculation into a picture of two sliding apertures.

The core insight

For an incoherent imaging system the OTF is the normalized autocorrelation of the pupil: the MTF at spatial frequency f equals the fractional overlap area of the aperture with a copy of itself shifted by λ·d_i·f. Two apertures fully overlap at f = 0 (MTF = 1) and separate completely at the cutoff (MTF = 0).

Where this lives today

Every lens datasheet is an MTF chart: contrast versus cycles/mm at several field positions and orientations, because that curve — not a lone resolution figure — predicts perceived sharpness. In incoherent fluorescence microscopy the OTF cutoff $2\,\mathrm{NA}/\lambda$ sets the finest transferable detail, and deconvolution divides the measured spectrum by the OTF to sharpen images; structured-illumination microscopy deliberately folds high object frequencies down into the passband to reach beyond the classical cutoff. Semiconductor photolithography lives at the cutoff, using off-axis illumination, phase-shift masks and apodized pupils to reshape the effective transfer function and print features finer than a naive $\lambda/\mathrm{NA}$ estimate (the partially coherent regime interpolating between the coherent ATF and incoherent OTF). In astronomy the long-exposure OTF is the telescope's own OTF times the atmosphere's MTF (set by the Fried parameter $r_0$ ); adaptive optics restores the pupil phase to recover the diffraction-limited OTF. The O11 photonics article on Imaging Systems and Optical Transfer Functions revisits these instruments in depth — this article owns the underlying theory.

The physical picture

Photographers and lithographers do not quote a single "resolution" number, because it hides the story. Instead they quote a curve: how much contrast a sinusoidal target retains as its bars are packed tighter and tighter. That curve is the MTF, and it is the most honest one-line summary of an imaging system. Two lenses can share a cutoff yet look utterly different — the one with high mid-frequency MTF renders "punchy," the one that merely resolves fine lines at feeble contrast renders "soft." The engine underneath is Fourier duality. A wide pupil supports high spatial frequencies and makes a tight point-spread function; a small pupil truncates them and makes a broad one. Because incoherent intensities add (not amplitudes), the image intensity is the ideal image convolved with the intensity point-spread function, and the transfer function is the pupil autocorrelated with itself. The autocorrelation of something of width D has width 2D — which is exactly why the incoherent system reaches twice the coherent cutoff frequency, though with a gently declining rather than brick-wall response. The punchline figure makes this quantitative for a real lens. The second figure shows the operational meaning: feed in a grating of Michelson contrast C, and the same grating emerges with contrast MTF·C — visibly greyer at high frequency.

Figure

The diffraction-limited MTF of an f/2 lens — real numbers on the axis: incoherent cutoff 909 lp/mm, coherent cutoff at exactly half

Figure

Operational meaning of MTF — a grating of contrast C emerges with contrast MTF·C

The core equations

For incoherent light the image intensity is the ideal (geometric) image intensity convolved with the intensity point-spread function. We assume unit magnification and an isoplanatic patch (space-invariant blur), so a single kernel describes the whole field.

g_i(x,y)=\iint g_g(x',y')\,|h(x-x',y-y')|^2\,\mathrm{d}x'\,\mathrm{d}y' \qquad (1)

where: $g_i$ — image intensity; $g_g$ — ideal geometric image intensity; $|h|^2$ — intensity point-spread function (the incoherent impulse response); $(x,y)$ — image-plane coordinates. Reading it physically: every point of the ideal image is replaced by a little glowing PSF-shaped blob, and the blobs sum. Fourier-transforming the convolution turns it into a product. Normalizing to unity at zero frequency defines the OTF, guaranteeing that a uniform object images to a uniform intensity.

\mathcal{H}(f_X,f_Y)=\frac{\mathcal{F}\{|h|^2\}}{\mathcal{F}\{|h|^2\}\big|_{0,0}},\qquad \mathcal{H}=\mathrm{MTF}\;e^{\,i\,\mathrm{PTF}} \qquad (2)

where: $\mathcal{H}$ — optical transfer function; $\mathcal{F}$ — 2-D Fourier transform; $(f_X,f_Y)$ — spatial frequencies (cycles per unit image-plane length); $\mathrm{MTF}=|\mathcal{H}|$ — modulation transfer function; $\mathrm{PTF}=\arg\mathcal{H}$ — phase transfer function. The MTF answers "how much contrast survives?"; the PTF answers "did features shift or reverse?" The central structural result: for a diffraction-limited system the OTF is the normalized autocorrelation of the pupil. The shift variable is $\lambda d_i f$ , so spatial frequency literally maps to a displacement of one pupil across another.

\mathcal{H}(f_X,f_Y)=\frac{\displaystyle\iint P(x,y)\,P^{*}\!\big(x-\lambda d_i f_X,\,y-\lambda d_i f_Y\big)\,\mathrm{d}x\,\mathrm{d}y}{\displaystyle\iint |P(x,y)|^2\,\mathrm{d}x\,\mathrm{d}y} \qquad (3)

where: $P$ — complex exit-pupil function (aperture amplitude and any wavefront-error phase); $\lambda$ — wavelength; $d_i$ — exit-pupil-to-image distance. The denominator normalizes $\mathcal{H}(0)=1$ ; the numerator at shift $\lambda d_i f$ is the overlap area of two apertures. The coherent (amplitude) system and the incoherent (intensity) system have cutoffs differing by exactly two, because the amplitude transfer function is the pupil itself while the OTF is its autocorrelation.

f_c^{\text{coh}}=\frac{D}{2\lambda d_i}=\frac{\mathrm{NA}}{\lambda},\qquad f_c^{\text{incoh}}=\frac{D}{\lambda d_i}=\frac{2\,\mathrm{NA}}{\lambda} \qquad (4)

where: $D$ — exit-pupil diameter; $\mathrm{NA}\approx D/(2 d_i)$ — image-space numerical aperture. For a clear circular pupil the overlap area has a closed form — the classic "two-circle lens" area — giving the diffraction-limited MTF.

\mathrm{MTF}(f)=\frac{2}{\pi}\left[\arccos\!\Big(\frac{f}{f_c}\Big)-\frac{f}{f_c}\sqrt{1-\Big(\frac{f}{f_c}\Big)^2}\;\right],\quad 0\le f\le f_c \qquad (5)

where: $f$ — radial spatial frequency; $f_c=2\,\mathrm{NA}/\lambda$ — incoherent cutoff. This gives $\mathrm{MTF}(0)=1$ , $\mathrm{MTF}(f_c)=0$ , and the workhorse anchor $\mathrm{MTF}(f_c/2)=0.391$ .

Where the OTF comes from

Derivation AFrom the point-spread function to the transfer function

The goal is to show that an incoherent, shift-invariant system does nothing more exotic than multiply each spatial-frequency component of the ideal image by a single complex number — the OTF — so that "imaging" becomes "filtering." The strategy is to Fourier-transform the intensity-convolution law (1) and read off the multiplier, then fix its normalization so that flat objects stay flat. Watch how the impulse response and the transfer function turn out to be a Fourier-transform pair, the master relationship of the whole subject.

Step 1. Write the incoherent imaging law as a convolution, $g_i = g_g * |h|^2$ . Why a convolution and not something worse: linearity lets us build any object from point sources, and shift-invariance says each point images to the same blob merely re-centred — the definition of convolution.

Step 2. Apply the convolution theorem, $\mathcal{F}\{g_g * |h|^2\} = \tilde g_g \cdot \mathcal{F}\{|h|^2\}$ . Why this is the whole trick: convolution in space is ordinary multiplication in frequency, so the tangled integral becomes a frequency-by-frequency scalar multiply.

Step 3. Normalize by the zero-frequency value, $\mathcal{H} = \mathcal{F}\{|h|^2\}/\mathcal{F}\{|h|^2\}|_0$ . Why divide: without it a bright uniform scene would change average brightness on imaging; forcing $\mathcal{H}(0)=1$ conserves total flux and makes MTF a pure contrast ratio.

Step 4. Read off the meaning: each Fourier component of the ideal image is scaled by $|\mathcal{H}| = \mathrm{MTF}$ and phase-shifted by $\arg\mathcal{H} = \mathrm{PTF}$ . Why this matters: it reduces "what does this lens do to any scene?" to one measured curve, exactly as a filter's frequency response summarizes a circuit. *Verified: $\mathcal{H}(0)=1$ by construction; Fourier-pair PSF↔OTF consistent with Goodman ch. 6 and Saleh & Teich §4.4 — verify_optical-transfer-function.py.*

Go deeper · Derivation B · 1 minThe circular-aperture MTF as sliding-aperture overlap

Derivation BThe circular-aperture MTF as sliding-aperture overlap

The goal is the closed-form diffraction-limited MTF (5) and its cutoff, obtained without a single diffraction integral — purely as the geometric overlap of two identical circular pupils sliding across each other. The strategy chains two facts: the incoherent OTF is the pupil autocorrelation (equation 3), and the autocorrelation of a uniform disc is its self-overlap area. The reward is a picture you can draw on a napkin that yields the exact curve optical designers plot.

Step 1. For a diffraction-limited system the amplitude PSF is the Fraunhofer pattern of the pupil, so its transform — the amplitude transfer function — is the scaled pupil $P$ itself. Squaring the PSF and transforming gives $\mathcal{F}\{|h|^2\}=P\star P$ , the autocorrelation. Why squaring becomes autocorrelation: the transform of a modulus-square is the self-correlation of the transform — the autocorrelation theorem.

Step 2. Take two discs of radius $a=D/2$ whose centres are separated by $s=\lambda d_i f$ . Their overlap area is the standard symmetric-lens formula $A = 2a^2\arccos(s/2a) - \tfrac{s}{2}\sqrt{4a^2-s^2}$ . Why this shape: the intersection of two equal circles is a lens (vesica) bounded by two arcs; its area is elementary integral geometry.

Step 3. Identify the shift ratio with normalized frequency, $s/2a = \lambda d_i f/D = f/f_c$ with $f_c = D/\lambda d_i$ . Why the cutoff appears here: overlap vanishes when the discs just touch, $s = 2a$ , i.e. $f = f_c$ — the aperture literally runs out of itself.

Step 4. Divide the overlap by the single-disc area $\pi a^2$ to normalize, giving equation (5). Endpoints: $\mathrm{MTF}(0)=1$ (full overlap), $\mathrm{MTF}(f_c)=0$ (no overlap), and $\mathrm{MTF}(f_c/2)=\tfrac{2}{\pi}[\arccos\tfrac12 - \tfrac12\sqrt{3/4}] = 0.391$ . Why 0.391 is worth memorizing: it is the diffraction-limited contrast at half-cutoff, the reference every real lens is judged against. *Verified: MTF(0)=1, MTF(1·f_c)=0, MTF(0.5 f_c)=0.3910, MTF(0.33 f_c)=0.588; slit-pupil autocorrelation is the triangle MTF(½)=0.5 — verify_optical-transfer-function.py.*

Go deeper · Derivation C · 1 minWhy the incoherent cutoff is exactly twice the coherent cutoff

Derivation CWhy the incoherent cutoff is exactly twice the coherent cutoff

The goal is to nail the factor of two that trips up every first course: coherent imaging cuts off at $\mathrm{NA}/\lambda$ , incoherent at $2\,\mathrm{NA}/\lambda$ , yet coherent imaging is not "twice as bad." The strategy is a support-counting argument — track the frequency extent of the amplitude transfer function versus its autocorrelation — and then separate the reach of a system from the quality of its response. This is the cleanest place to see that "resolution" is not one number.

Step 1. The amplitude (coherent) transfer function is $H(f)=P(\lambda d_i f)$ , nonzero only where $|\lambda d_i f|\le D/2$ , i.e. up to $f_c^{\text{coh}}=D/2\lambda d_i=\mathrm{NA}/\lambda$ . Why the pupil is the filter: under coherent light the system is linear in complex field, and a diffraction-limited lens passes exactly the angular spectrum its aperture admits — a brick wall.

Step 2. The incoherent OTF is the autocorrelation $H\star H$ . A function supported on $|f|\le a$ has an autocorrelation supported on $|f|\le 2a$ . Why the doubling: correlation sums the reach of both copies — the last overlap survives until one aperture is shifted by a full diameter.

Step 3. Hence $f_c^{\text{incoh}}=2 f_c^{\text{coh}}=2\,\mathrm{NA}/\lambda$ , but with a smoothly declining response, not a brick wall. Why coherent is not simply better: its brick-wall filter rings — sharp edges show overshoot, and speckle and edge artefacts appear — so the two regimes differ in character, not by a scale factor. *Verified: coherent 455 lp/mm, incoherent 909 lp/mm for the f/2, 550 nm case, ratio exactly 2 — verify_optical-transfer-function.py.*

Limiting cases

Case	Result	Meaning
Zero frequency	$\mathrm{OTF}(0)=1$	Normalization: the uniform-brightness component always passes perfectly; flux conserved.
At the cutoff	$\mathrm{MTF}(f)=0$ for $f\ge f_c=2\,\mathrm{NA}/\lambda$	No incoherent system transfers detail finer than the cutoff — hard zero beyond it.
Coherent vs incoherent	$f_c^{\text{incoh}}=2\,f_c^{\text{coh}}$	Autocorrelating the pupil doubles its frequency support.
Symmetric (aberration-free) PSF	$\mathrm{PTF}\in\{0,\pi\}$	A real, even PSF gives a real OTF: phase is a sign only.
Strong defocus / aberration	$\mathrm{MTF}(f)<0$ possible	Zeros and negative lobes appear: contrast reversal ("spurious resolution").

Common trapCoherent and incoherent "cutoff frequencies" are not the same quantity — and the coherent system is not simply "twice as bad"

The amplitude (coherent) transfer function cuts off at $\mathrm{NA}/\lambda$ ; the incoherent OTF reaches $2\,\mathrm{NA}/\lambda$ , a factor of two higher. It is tempting to conclude "coherent imaging has half the resolution," but that misreads the physics. The coherent transfer function is a brick-wall filter linear in amplitude: it produces edge ringing, speckle and interference artefacts, and its "resolution" depends on object phase. The incoherent OTF is a soft, monotone response linear in intensity. They are qualitatively different mappings — do not relate their MTFs by a trivial rescaling, and never mix an amplitude-domain cutoff with an intensity-domain one in the same formula.

Worked examples

1 · The MTF curve of a real f/2 camera lens (confidence builder)

A diffraction-limited f/2 lens images at $\lambda = 550$ nm (green). Find its incoherent cutoff, its contrast at 300 lp/mm, and its half-cutoff contrast.

Step 1. The image-space numerical aperture of an f/2 cone is $\mathrm{NA} = 1/(2\cdot f/\#) = 1/(2\cdot 2) = 0.25$ .

Step 2. Incoherent cutoff $f_c = 2\,\mathrm{NA}/\lambda = 2(0.25)/(0.550\ \mu\mathrm{m}) = 0.909\ \mathrm{cycles}/\mu\mathrm{m} = \mathbf{909\ lp/mm}$ . The coherent cutoff is exactly half, 455 lp/mm.

Step 3. At 300 lp/mm the normalized frequency is $\hat f = 300/909 = 0.330$ , so $\mathrm{MTF} = \tfrac{2}{\pi}[\arccos 0.330 - 0.330\sqrt{1-0.330^2}] = \mathbf{0.59}$ — nearly 60% contrast survives at 300 lp/mm.

Step 4. At half-cutoff (455 lp/mm) the diffraction-limited value is the memorizable $\mathrm{MTF}(f_c/2) = \mathbf{0.39}$ . A real lens sits below this curve; a lens on it is "diffraction-limited."

2 · Reading the Abbe limit off the OTF (exam level)

Show that the incoherent OTF cutoff reproduces the Abbe resolution limit, and quote the finest resolvable period for the lens above.

Step 1. The finest transferable spatial frequency is $f_c = 2\,\mathrm{NA}/\lambda$ , so the finest transferable period is $1/f_c = \lambda/(2\,\mathrm{NA})$ — precisely the Abbe resolution limit.

Step 2. For $\mathrm{NA} = 0.25$ , $\lambda = 550$ nm: smallest period $= 0.550/(2\cdot 0.25) = \mathbf{1.10\ \mu m}$ (i.e. 909 lp/mm), matching the cutoff from Example 1.

Step 3. The cross-check with the point-spread function: the Airy-disk radius and the OTF cutoff are Fourier partners, so this same NA fixes both the real-space blur (see Point Spread Function) and the frequency-space cutoff — one number, two faces.

Physics of the portrait

The OTF is not named after a person; it is a method, and its story is the story of importing communication-engineering thinking into optics. Before the 1940s, lens quality was argued over with "lines per millimetre" resolution charts — a single threshold number. The reframing that an imaging system is a linear filter with a frequency response, so that the right description is a whole transfer curve, is credited to Pierre-Michel Duffieux, whose 1946 monograph L'Intégrale de Fourier et ses applications à l'optique founded Fourier optics. The concept was sharpened into the OTF/MTF vocabulary through the 1950s–60s (Hopkins, Linfoot, and the systems-optics community), and Goodman's textbook made it canonical.

Origins · 1 minoptical transfer function and how this was discovered

optical transfer function

method · Duffieux 1946 → modern Fourier optics · the frequency-response view of imaging

The OTF crystallized the idea that a lens is a filter. Pierre-Michel Duffieux (1891–1976), working in Besançon, published L'Intégrale de Fourier et ses applications à l'optique in 1946 — the book that recast image formation as a Fourier-domain operation and let optics borrow the linear-systems machinery already mature in electronics. Two payoffs followed at once. First, "resolution" dissolved into a curve: the MTF quantifies contrast at every spatial frequency, so a lens that resolves fine lines faintly and one that renders coarse detail crisply are no longer forced onto a single number. Second, the incoherent OTF revealed itself as the pupil's own autocorrelation, geometrizing diffraction as two sliding apertures and explaining the factor-of-two cutoff with a picture. The framework now underwrites lens specification, microscope deconvolution, adaptive-optics metrics, and the resolution-enhancement tricks of photolithography.

History

1873

— Ernst Abbe frames image formation in terms of diffraction orders captured by the aperture — the physical seed of a frequency-domain view

1934

— Frits Zernike introduces phase contrast, manipulating the object spectrum in the focal plane

1946

— Pierre-Michel Duffieux publishes the Fourier-optics monograph founding the transfer-function description of imaging

1955

— H. H. Hopkins formalizes partial coherence and the transmission cross-coefficient for imaging

1968

— Joseph Goodman's Introduction to Fourier Optics makes the OTF/MTF the standard language of imaging systems

Going further

Texts. Goodman, Introduction to Fourier Optics, 3rd ed., ch. 6 (frequency analysis of imaging) · Saleh & Teich, Fundamentals of Photonics, 3rd ed., §4.4 (transfer functions) · Hecht, Optics, 5th ed., §11.3, §13.2 · Born & Wolf, Principles of Optics, ch. 9.

Provenance (base-corpus migration). The base derivation and numbers are standard and were re-verified from scratch: OTF = normalized pupil autocorrelation and the cutoff doubling (Goodman ch. 6); the circular-aperture overlap formula with $\mathrm{MTF}(0)=1$ , $\mathrm{MTF}(f_c)=0$ , $\mathrm{MTF}(0.5f_c)=0.391$ ; $f_c^{\text{incoh}}=2\,\mathrm{NA}/\lambda$ , $f_c^{\text{coh}}=\mathrm{NA}/\lambda$ ; and the Abbe cross-check $\lambda/(2\,\mathrm{NA})$ . No base error found — every quoted value reproduced by verify_optical-transfer-function.py (16 checks).

Builds on: Fraunhofer Diffraction · Circular Aperture Diffraction and the Airy Pattern · Numerical Aperture and f-number. Leads to: Point Spread Function · Rayleigh Resolution Criterion · Spatial Filtering and the 4f System · Imaging Systems and Optical Transfer Functions (O11).

Is the OTF the same thing as the MTF?

No — the MTF is the modulus of the OTF. The full OTF is complex: $\mathcal{H} = \mathrm{MTF}\,e^{i\,\mathrm{PTF}}$ . The MTF (how much contrast survives) is what datasheets usually plot, but the phase transfer function matters whenever the PSF is asymmetric (coma, decentring), because it shifts features sideways and can distort edges. For a symmetric, aberration-free system the PTF is just 0 or π, so the OTF is real and the MTF plus a sign carries everything.

Why normalize the OTF to 1 at zero frequency?

The zero-frequency (DC) component is the average brightness. If the transfer function did not equal 1 there, a uniform grey object would come out brighter or darker on imaging — unphysical for a passive lens that only redistributes light. Normalizing forces flux conservation and makes the MTF a pure contrast ratio between 0 and 1, so "MTF = 0.4" means "40% of the object contrast survives," independent of exposure.

How is the OTF related to the point-spread function?

They are a Fourier-transform pair. The intensity PSF (the image of a point) transforms into the OTF; equivalently, the OTF is the normalized autocorrelation of the pupil while the PSF is the squared Fourier transform of the pupil. Everything about resolution can be read in either domain: a tight PSF ⇔ a broad, high MTF; a fat PSF ⇔ a narrow MTF. The companion article Point Spread Function develops the real-space side.

Can a system have an MTF above 1 at some frequency?

Not a passive imaging system — the MTF is bounded by 1 because it is a normalized overlap (or contrast ratio), and a lens cannot create contrast. Apparent "peaking" above 1 shows up only when you post-process (sharpening, deconvolution, or an electronic edge-enhancement filter), which multiplies the measured spectrum by a gain. That is a separate stage bolted onto the optics; the optical transfer function itself never exceeds unity.

What is "spurious resolution" on a defocused test chart?

Defocus adds a quadratic phase across the pupil. Autocorrelating that complex pupil produces an OTF that dips to zero at some intermediate frequency and then goes negative beyond it. A negative OTF is a contrast reversal — bars that were black now image white — so a defocused lens can appear to "resolve" a fine pattern, but with the wrong polarity. It is a real, repeatable artefact, not noise, and it is the classic warning that "resolves it" is not the same as "images it faithfully."

How do the OTFs of a lens, a sensor, and the atmosphere combine?

They multiply, frequency by frequency: $\mathrm{MTF}_{\text{system}} = \mathrm{MTF}_{\text{lens}}\times\mathrm{MTF}_{\text{sensor}}\times\mathrm{MTF}_{\text{atmosphere}}\times\cdots$ , provided each stage is linear and shift-invariant and the light is incoherent between stages. This cascade rule is why a superb lens on a coarse sensor still looks soft, and why the weakest link dominates. It also warns you that adding stages can only lower the MTF at any given frequency — never raise it.