Fourier-Transform Spectroscopy

Fourier-transform spectroscopy turns the ordinary spectrometer inside out. A grating or prism fans the colours across space and reads them channel by channel; a Fourier-transform spectrometer refuses to separate the colours at all. Instead it splits a single beam in a scanning Michelson interferometer, lets the two copies interfere, and records one number — the detected brightness — as the path difference is swept. The whole spectrum is folded into that single scan, the interferogram, and a Fourier transform unfolds it back into colour. No slit, no ruling, no moving grating: just one mirror sliding through a few centimetres while a detector watches the fringes bloom and fade. The reason this works is the two-beam interference term. When a field is split, delayed by a time $\tau$ , and recombined, the modulated part of the output is the field's own autocorrelation $\langle E^*(t)E(t+\tau)\rangle$ — the very quantity that measures temporal coherence. The Wiener–Khinchin theorem then supplies the punchline: the Fourier transform of that autocorrelation is the power spectrum. So the interferogram is not a picture of the spectrum but its Fourier transform, and reading one off the other is a single numerical step. This one idea makes the infrared spectrum of almost every molecule on Earth accessible in seconds, and it carries two structural bonuses — the multiplex and throughput advantages — that a slit spectrometer can never match.

Key takeaways

Fourier-transform spectroscopy (FTS) recovers a spectrum $S(\nu)$ not by dispersing light in space but by recording an interferogram: the output of a scanning Michelson interferometer versus optical path difference $x$ .
The modulated interferogram is the real part of the field autocorrelation $\Gamma(\tau)$ , $\tau=x/c$ — the same degree of temporal coherence that governs fringe visibility.
By the Wiener–Khinchin theorem, the Fourier transform of that autocorrelation is the power spectrum, so $S(\tilde\nu)\propto\int I_{\text{AC}}(x)\cos(2\pi\tilde\nu x)\,dx$ — a delay scan replaces a grating and every colour is measured at once.
The maximum path difference sets the resolution, $\delta\tilde\nu=1/(2\Delta_{\max})$ ; the sampling step sets the Nyquist limit against aliasing.
The multiplex (Fellgett) and throughput (Jacquinot) advantages give FTS a large signal-to-noise edge in the infrared.
Key idea: the interferogram of a scanning Michelson is the field's autocorrelation; its Fourier transform with respect to path delay is the source's power spectrum, and the scan length alone fixes the resolution at δσ = 1/(2Δmax).

The core insight

A two-beam interferometer measures the autocorrelation of the light, not the spectrum. Sweep the path difference to record the interferogram $I(x)$ ; subtract its mean and Fourier-transform it, and the Wiener–Khinchin theorem returns the spectrum $S(\tilde\nu)$ . Because the transform pair is autocorrelation ↔ spectrum, narrow lines make long interferograms and broad bands make short bursts — and the longest delay you reach, $\Delta_{\max}$ , is exactly what limits how finely you can resolve.

Where this lives today

Fourier-transform spectroscopy is the default way to take an infrared spectrum. Virtually every chemistry lab runs an FTIR: a broadband source, a Michelson with a potassium-bromide or diamond beamsplitter, and a helium–neon reference laser deliver a molecule's vibrational fingerprint across roughly 400–4000 cm⁻¹ in seconds, with the Fellgett and Jacquinot advantages supplying the sensitivity to see thin films, trace gases, and reactions as they run. The same architecture, scaled up, flies in space: atmospheric sounders such as IASI and climate instruments measuring CO₂, CH₄, and H₂O rely on large maximum path differences for the fine resolution needed to pick apart rotational–vibrational lines. The idea also generalises far beyond optics. Fourier-transform nuclear magnetic resonance and Fourier-transform ion-cyclotron-resonance mass spectrometry record a time-domain signal and transform it to a spectrum by exactly the same multiplex logic, inheriting the same window-length and Nyquist rules. And because the interferogram envelope is the degree of temporal coherence, an FTS scan doubles as a coherence-length meter: a Lorentzian decay reveals lifetime or collision broadening, a Gaussian decay reveals Doppler broadening — the line-broadening mechanism read straight off the fading of the fringes.

The physical picture

Imagine building a spectrum out of pure tones. Each optical frequency $\nu$ , passing through a two-beam interferometer, writes a single cosine $\cos(2\pi\nu x/c)$ into the path-difference axis. A real source is a chorus of such tones with weights $S(\nu)$ ; the detector adds all their cosines together, and the result is the interferogram. Reverse the operation — a Fourier transform — and the weights fall back out. The interferometer is a physical Fourier synthesiser, and the computer that transforms its output is the analyser that reads the score back. At zero path difference every tone is in step and they pile up into a tall central burst. As the mirror moves, the tones drift out of phase and the fringes wash out — quickly for a broad spectrum, slowly for a narrow one. That decay is the loss of temporal coherence, and its width in delay is the coherence length. Narrow in frequency means wide in delay, and vice versa: the reciprocity that runs through all of Fourier optics.

Figure

Interferogram → spectrum for three sources: a single line gives an undamped cosine (a δ-spike), a close doublet gives a beating cosine (two lines), and a broad band gives a sharp decaying burst (a wide hump). The Fourier transform reads one from the other.

A reference laser makes all of this metrology rather than art. A helium–neon beam runs through the same interferometer and marks off equal steps of path difference by its own single fringe; every sample of the interferogram is therefore taken at a path difference known to a fraction of a wavelength. FTS inherits its wavenumber accuracy directly from that reference wavelength — which is why the method is prized not only for sensitivity but for the sheer trustworthiness of its frequency axis. The single most important number in the whole method is how far the mirror travels. The finite scan truncates the interferogram, and truncation blurs the recovered spectrum. The next figure makes the rule concrete: resolution and scan length are simple reciprocals.

Figure

The resolution punchline: wavenumber resolution δσ = 1/(2Δmax) falls as the reciprocal of the maximum path difference. A 10 cm scan resolves 0.05 cm⁻¹ (equivalently 1.5 GHz) — marked on the curve.

The core equations

Start from what the detector actually sees. A 50:50 beamsplitter makes two copies of the field, one delayed by $\tau=x/c$ , and the detector time-averages the squared sum. The fast optical oscillation averages away, leaving a constant plus one modulated term — and that term is the field talking to a delayed copy of itself.

I(x)=\Gamma(0)+\operatorname{Re}\Gamma(\tau),\qquad \Gamma(\tau)=\langle E^*(t)\,E(t+\tau)\rangle,\qquad \tau=\frac{x}{c} \qquad (1)

where: $I(x)$ — detected intensity at optical path difference $x$ ; $\Gamma(\tau)$ — the field autocorrelation (self-coherence) function; $\Gamma(0)$ — the mean intensity (the DC level); $\tau$ — the time delay between the two arms. The measurement is an autocorrelation of the light, full stop; the constant $\Gamma(0)$ carries no colour information and is subtracted before transforming. The bridge from that autocorrelation to the spectrum is the Wiener–Khinchin theorem. For a stationary field the autocorrelation and the power spectral density are a Fourier-transform pair — the spectrum is not measured, it is computed.

\Gamma(\tau)=\int_0^{\infty} S(\nu)\,e^{-i2\pi\nu\tau}\,d\nu \;\;\Longleftrightarrow\;\; S(\nu)=\int_{-\infty}^{\infty}\Gamma(\tau)\,e^{\,i2\pi\nu\tau}\,d\tau \qquad (2)

where: $S(\nu)$ — power spectral density (the source spectrum); $\nu$ — optical frequency. Reading left to right builds the interferogram from the spectrum; reading right to left is what the spectrometer's computer does. Everything else is bookkeeping on this one pair. In practice the field is real, the spectrum one-sided, and the interferogram symmetric, so the complex transform collapses to a cosine transform in the directly measured path difference $x$ . Working in wavenumber $\tilde\nu=\nu/c=1/\lambda$ (the spectroscopist's unit, cm $^{-1}$ ) removes the factor of $c$ entirely.

S(\tilde\nu)\;\propto\;\int_{-\infty}^{\infty}\big[I(x)-\bar I\,\big]\,\cos(2\pi\tilde\nu\,x)\,dx \qquad (3)

where: $\tilde\nu$ — wavenumber; $\bar I=\Gamma(0)$ — the subtracted mean level; $x$ — optical path difference, twice the mirror displacement. The cosine transform is its own inverse up to a constant, so the same operation runs in both directions. The scan cannot be infinite, and the truncation is what limits resolution. Cutting the interferogram off at $|x|\le\Delta_{\max}$ convolves the true spectrum with a sinc instrument line shape whose width sets the smallest resolvable interval.

\delta\tilde\nu=\frac{1}{2\Delta_{\max}},\qquad \delta\nu=\frac{c}{2\Delta_{\max}} \qquad (4)

where: $\delta\tilde\nu$ — wavenumber resolution (cm $^{-1}$ for $\Delta_{\max}$ in cm); $\delta\nu$ — frequency resolution; $\Delta_{\max}$ — maximum optical path difference. Resolution is pure geometry — how far the mirror travels — and owes nothing to the detector. Sampling at steps $\Delta x$ then sets the Nyquist ceiling $\tilde\nu_{\max}=1/(2\Delta x)$ above which frequencies alias. Finally, the reason FTS conquered the infrared: because every spectral element is measured throughout the whole scan rather than one at a time, a detector-noise-limited FTS beats a scanning monochromator by the square root of the number of elements.

\left(\frac{S}{N}\right)_{\text{FTS}}=\sqrt{M}\,\left(\frac{S}{N}\right)_{\text{disp}} \qquad (5)

where: $M$ — number of resolved spectral elements measured simultaneously; the equality is the Fellgett (multiplex) advantage and holds when additive detector noise dominates.

Derivation AFrom the interferogram to the spectrum by Wiener–Khinchin

The goal is to show, from scratch, why a scanning two-beam interferometer measures a Fourier transform of the spectrum rather than the spectrum itself, and therefore why a single numerical transform of the recorded interferogram returns the colours. We build the detected intensity from the superposed fields, isolate the modulated part as the field autocorrelation, and invoke the Wiener–Khinchin theorem to identify its transform as the power spectrum — the exact chain equations (1)–(3) assert.

Step 1. Split the field into two equal copies and delay one by $\tau=x/c$ , so the recombined field is $E_{\text{out}}(t)=\tfrac{1}{\sqrt2}[E(t)+E(t+\tau)]$ ; the detector reports the time average $I=\langle|E_{\text{out}}|^2\rangle=\Gamma(0)+\operatorname{Re}\langle E^*(t)E(t+\tau)\rangle$ . Why: a square-law detector integrates over millions of optical cycles, so every term oscillating at the optical frequency averages to zero except the correlation between the field and its delayed self.

Step 2. Recognise the surviving term as the autocorrelation $\Gamma(\tau)=\langle E^*(t)E(t+\tau)\rangle$ , so the modulated interferogram is exactly $I_{\text{AC}}(x)=\operatorname{Re}\Gamma(x/c)$ . Why: this is the same self-coherence function that sets fringe visibility, so the interferogram literally traces the light's degree of temporal coherence as the delay grows.

Step 3. Expand the field in its spectral components, $E(t)=\int\tilde E(\nu)e^{-i2\pi\nu t}\,d\nu$ , and use stationarity, $\langle\tilde E^*(\nu)\tilde E(\nu')\rangle=S(\nu)\,\delta(\nu-\nu')$ , to collapse the double integral to $\Gamma(\tau)=\int_0^\infty S(\nu)e^{-i2\pi\nu\tau}\,d\nu$ . Why: a stationary source cannot correlate different frequencies, so the delta function reduces the correlation of the whole field to a single integral weighted by the spectrum — this is the Wiener–Khinchin theorem in the making.

Step 4. Take the real part and invert the cosine transform: $\operatorname{Re}\Gamma(\tau)=\int_0^\infty S(\nu)\cos(2\pi\nu\tau)\,d\nu$ , so $S(\tilde\nu)\propto\int I_{\text{AC}}(x)\cos(2\pi\tilde\nu x)\,dx$ with $\tau=x/c$ . Why: the cosine transform is self-inverse up to a constant, so the very operation that built the interferogram from the spectrum, run backwards on the measured interferogram, hands back the spectrum. *Verified: for a monochromatic input $S(\nu)=S_0\delta(\nu-\nu_0)$ the interferogram reduces to $1+\cos(2\pi\nu_0 x/c)$ , and a numerical cosine transform of a two-line (doublet) spectrum reproduces the beating interferogram and recovers both line positions, in verify_aq-optics-fourier-transform-spectroscopy.py.*

Go deeper · Derivation B · 1 minWhy the scan length fixes the resolution

Derivation BWhy the scan length fixes the resolution

The aim here is to pin down the single most consequential specification of a Fourier-transform spectrometer — its resolving power — and to show it is set entirely by how far the mirror travels, not by the detector or the source. We treat the finite scan as a rectangular window on the interferogram and track what that truncation does to the recovered spectrum, arriving at $\delta\tilde\nu=1/(2\Delta_{\max})$ and the Nyquist sampling condition.

Step 1. Write the finite scan as the ideal (infinite) interferogram multiplied by a boxcar of half-width $\Delta_{\max}$ : $I_{\text{meas}}(x)=I(x)\cdot\operatorname{rect}(x/2\Delta_{\max})$ . Why: no instrument scans to infinite delay, and the boxcar captures exactly what is thrown away — everything beyond the last mirror position.

Step 2. Transforming a product gives a convolution, so the recovered spectrum is the true spectrum smeared by the transform of the boxcar, $S_{\text{meas}}(\tilde\nu)=S(\tilde\nu)*\text{ILS}(\tilde\nu)$ with $\text{ILS}(\tilde\nu)=2\Delta_{\max}\,\operatorname{sinc}(2\pi\Delta_{\max}\tilde\nu)$ . Why: truncation in delay must blur in frequency; the sharper you want the spectrum, the longer you must scan — the two are Fourier reciprocals.

Step 3. Read the width of that sinc: its first zero sits at $\tilde\nu=1/(2\Delta_{\max})$ , which is the smallest interval over which two lines stay distinct (the Rayleigh-like criterion), so $\delta\tilde\nu=1/(2\Delta_{\max})$ and $\delta\nu=c/(2\Delta_{\max})$ . Why: two lines closer than the first zero merge under the instrument line shape, so the sinc half-width is the resolution — a pure consequence of the scan length.

Step 4. Sample the interferogram at steps $\Delta x$ ; the recovered spectrum then repeats with period $1/\Delta x$ , so to avoid overlap the spectrum must lie below $\tilde\nu_{\max}=1/(2\Delta x)$ (equivalently $\Delta x\le\lambda_{\min}/2$ ). Why: discrete sampling is periodic in frequency, so anything above the Nyquist wavenumber folds back into the band — a reference laser fixes $\Delta x$ precisely to keep that ceiling under control. *Verified: $\delta\tilde\nu=1/(2\Delta_{\max})=0.05$ cm⁻¹ (equivalently $\delta\nu=1.5$ GHz) at $\Delta_{\max}=10$ cm, the $\Delta_{\max}\ge0.5$ cm needed to split lines 1 cm⁻¹ apart, and the Nyquist relation $\Delta x=\lambda_{\min}/2$ , in verify_aq-optics-fourier-transform-spectroscopy.py.*

Limiting cases

Case	Result	Meaning
Monochromatic source	$I(x)\propto 1+\cos(2\pi\tilde\nu_0 x)$	a single line gives an undamped cosine; its transform is a δ-spike at $\tilde\nu_0$
Lorentzian line, width $\Delta\tilde\nu$	$\|\gamma(\tau)\|=e^{-\pi c\Delta\tilde\nu\,\|\tau\|}$	exponential envelope; coherence length $\ell_c=1/(\pi\Delta\tilde\nu)$ read straight off the decay
Gaussian (Doppler) line	$\|\gamma(\tau)\|=\exp[-(\pi c\Delta\tilde\nu\,\tau)^2/4\ln 2]$	Gaussian envelope; the transform of a Gaussian is a Gaussian, recovering the Doppler profile
Finite scan $\Delta_{\max}$	$\delta\tilde\nu=1/(2\Delta_{\max})$	rectangular truncation → sinc instrument line shape with $\pm22\%$ sidelobes
Sampling step $\Delta x$	$\tilde\nu_{\max}=1/(2\Delta x)$	frequencies above the Nyquist limit alias; a bandpass or fine sampling prevents fold-over
Multiplex, $M$ elements	$(S/N)\times\sqrt{M}$	Fellgett gain in the detector-noise limit; vanishes (or reverses) when photon shot noise dominates

Common trap“A Fourier-transform spectrometer measures the spectrum directly.”

It does not. What the detector records is the interferogram — the autocorrelation of the field versus path difference — and the spectrum only appears after a Fourier transform. Two consequences follow that trip people up. First, a feature in the interferogram is not a spectral line: a single sharp line produces a cosine spread across the entire scan, and a broad band produces a narrow central burst — the opposite of intuition. Second, the resolution is set by how far the mirror travels, $\delta\tilde\nu=1/(2\Delta_{\max})$ , not by the detector, the source brightness, or the sampling density. If you want to split two close lines you must scan farther, and no amount of signal averaging at short delay will substitute. Reading the interferogram as if it were a spectrum, or expecting a better detector to buy resolution, are the two classic mistakes.

Worked examples

1 · Resolution of an FTIR spectrometer (confidence builder)

A benchtop FTIR scans to a maximum optical path difference of $\Delta_{\max}=10$ cm. What wavenumber resolution does it achieve, what is that in frequency, and how far must it scan to separate two lines 1 cm⁻¹ apart?

Step 1. Resolution is the reciprocal of twice the scan length: $\delta\tilde\nu=1/(2\Delta_{\max})=1/(2\times10\ \text{cm})=\mathbf{0.05\ cm^{-1}}$ . That is fine enough to resolve the rotational fine structure on a vibrational band of a small gas-phase molecule.

Step 2. Convert to frequency: $\delta\nu=c\,\delta\tilde\nu=(3.00\times10^{10}\ \text{cm/s})(0.05\ \text{cm}^{-1})=\mathbf{1.5\ GHz}$ . The same instrument therefore places optical features to better than a part in $10^{5}$ of a typical $10^{14}$ Hz optical frequency.

Step 3. Invert the resolution rule for the required scan: to reach $\delta\tilde\nu=1$ cm⁻¹ you need $\Delta_{\max}=1/(2\,\delta\tilde\nu)=1/(2\times1\ \text{cm}^{-1})=\mathbf{0.5\ cm}$ . Two lines 1 cm⁻¹ apart are just resolved by a mere half-centimetre of travel.

Step 4. The scaling is the design lever: quadrupling the wavenumber resolution to 0.0125 cm⁻¹ demands a $40$ cm scan — a metre-scale optical bench and correspondingly stringent mirror alignment. Resolution in FTS is bought with length, linearly. δσ = 0.05 cm⁻¹ (1.5 GHz) at 10 cm; 0.5 cm scan splits 1 cm⁻¹.

2 · The Fellgett multiplex advantage (exam level)

A mid-infrared spectrum is to be recorded across $M=1000$ resolution elements. Compare the signal-to-noise of a Fourier-transform spectrometer with a scanning grating monochromator that visits one element at a time, assuming the same total measurement time and detector-noise-limited operation.

Step 1. The monochromator spends only $1/M$ of the run on each element, gathering signal from one channel while the other $999$ go unmeasured. The FTS detector sees light from all elements throughout the entire scan — the multiplex principle.

Step 2. With additive detector noise independent of signal, the noise is the same for both instruments over equal time, while the FTS accumulates every element's photons the whole time. The signal-to-noise ratios then differ by the square root of the number of elements: $(S/N)_{\text{FTS}}=\sqrt{M}\,(S/N)_{\text{disp}}$ .

Step 3. Put in the number: $\sqrt{M}=\sqrt{1000}=\mathbf{31.6}$ . The Fourier-transform spectrometer delivers about a $32\times$ signal-to-noise improvement — or reaches the same quality in $1/1000$ of the time.

Step 4. The advantage is conditional. It is largest in the infrared, where cheap thermal detectors are detector-noise-limited; it shrinks toward the visible, and under photon shot noise from a bright source it can vanish or even reverse (the multiplex disadvantage), because piling all the light onto one detector piles up its shot noise too. √1000 ≈ 32× SNR, detector-noise-limited only.

Physics of the portrait

The whole method rests on one mathematical fact — that the autocorrelation of a signal and its power spectrum are a Fourier-transform pair — and that fact has a name.

Origins · 1 minWiener–Khinchin theorem and how this was discovered

Wiener–Khinchin theorem

Norbert Wiener 1930 · Aleksandr Khinchin 1934 · harmonic analysis of stationary processes

The Wiener–Khinchin theorem states that for a wide-sense-stationary random signal the power spectral density is the Fourier transform of the autocorrelation function. Norbert Wiener, the American founder of cybernetics, established it in his 1930 memoir Generalized Harmonic Analysis, giving a rigorous spectral theory for signals — like light from a thermal source or noise in a circuit — that have no ordinary Fourier transform because they never decay. The Soviet probabilist Aleksandr Khinchin gave the theorem its clean form for stationary stochastic processes in 1934, and the joint name has stuck. Its reach is enormous: it is why a noisy resistor has a definable spectrum (Johnson–Nyquist noise), why radio astronomers can build spectra from correlator outputs, and — the case here — why a two-beam interferometer, which physically measures the autocorrelation of an optical field, delivers the source's spectrum after one Fourier transform. The theorem is the mathematical licence for the entire enterprise of Fourier-transform spectroscopy: it guarantees that the interferogram and the spectrum carry identical information, just in conjugate variables.

History

1881

— Michelson builds his interferometer and later reads spectral structure from how fringe visibility decays with path difference

1892

— Rayleigh, in a letter to Michelson, notes that the visibility curve and the spectrum are Fourier related — and that visibility alone, lacking the phase, cannot fix an asymmetric line

1930

— Wiener publishes Generalized Harmonic Analysis, the autocorrelation–spectrum theorem for stationary signals

1934

— Khinchin casts the theorem in its standard stochastic-process form

1949

— Fellgett, in his thesis, recognises the multiplex (signal-to-noise) advantage of measuring all elements at once

1954

— Jacquinot quantifies the throughput (étendue) advantage of a circular aperture over a slit

1965

— Cooley and Tukey publish the fast Fourier transform, making routine numerical inversion of interferograms practical

1966

— Janine and Pierre Connes push astronomical Fourier-transform spectroscopy to very high resolution

Going further

References. Hecht, Optics, 5th ed., §7.4 and §12.1 (coherence, Fourier methods) · Loudon, The Quantum Theory of Light, 3rd ed., Ch. 3 (coherence and correlation functions) · Saleh & Teich, Fundamentals of Photonics, §11.3 (Fourier spectroscopy) · Demtröder, Laser Spectroscopy, Vol. 1, §4.4 (Fourier-transform spectroscopy) · Wiener (1930) and Khinchin (1934), the autocorrelation–spectrum theorem.

Provenance. Equations follow from the two-wave interference term and the Wiener–Khinchin theorem as in Hecht §12.1, Loudon Ch. 3, Saleh & Teich §11.3, and Demtröder §4.4. Recomputed and verified here: (i) the interferogram $I(x)=\Gamma(0)+\operatorname{Re}\Gamma(x/c)$ from superposition; (ii) the cosine-transform inversion $S(\tilde\nu)\propto\int I_{\text{AC}}(x)\cos(2\pi\tilde\nu x)\,dx$ , checked numerically on a doublet; (iii) resolution $\delta\tilde\nu=1/(2\Delta_{\max})=0.05$ cm⁻¹ (1.5 GHz) at $\Delta_{\max}=10$ cm; (iv) the Nyquist limit $\tilde\nu_{\max}=1/(2\Delta x)$ ; (v) the $\sqrt{M}$ Fellgett scaling in the detector-noise limit. Line-shape envelopes (Lorentzian $e^{-\pi c\Delta\tilde\nu|\tau|}$ , Gaussian) are standard Fourier pairs; no non-standard numerical constants are asserted.

Builds on. Two-Wave Interference and the Interference Term · Temporal Coherence and Coherence Length · Mutual Coherence Function and Degree of Coherence · Optical Path Difference and Phase Retardation · Fringe Visibility and Contrast. Leads to. Wavefront-Splitting and Amplitude-Splitting Interferometers · Spatial Coherence and the Van Cittert–Zernike Theorem · Conditions for Observable Interference.

Is the interferogram the same thing as the spectrum, just drawn differently?

No — they are Fourier transforms of each other, which means they carry the same information in conjugate variables but look nothing alike. The horizontal axis of the interferogram is optical path difference (a length), while the spectrum's axis is frequency or wavenumber. A single spectral line, sharp in frequency, spreads into a cosine that fills the whole interferogram; a broad band, wide in frequency, collapses into a narrow burst near zero path difference. You cannot read a spectral feature off the interferogram by eye; the transform is essential.

Why does scanning the mirror farther improve resolution?

Because truncating the interferogram at a maximum path difference $\Delta_{\max}$ is mathematically a rectangular window, and windowing in delay convolves the recovered spectrum with a sinc whose width is $1/(2\Delta_{\max})$ . Two lines closer than that width blur into one. The farther you scan, the narrower the sinc, and the finer the detail you can separate — so resolution and scan length are exact reciprocals, $\delta\tilde\nu=1/(2\Delta_{\max})$ . Nothing about the detector or the source changes this; only geometry does.

What is the reference laser for?

It supplies the ruler for the path-difference axis. A helium–neon laser passes through the same interferometer, and its single sharp fringe marks off equal, precisely known increments of path difference; the interferogram is sampled at each of those marks. This does two jobs: it fixes the sampling step $\Delta x$ (and hence the Nyquist limit) to a fraction of the laser wavelength, and it pins the absolute wavenumber scale of the final spectrum to the reference wavelength. FTS owes its exceptional frequency accuracy almost entirely to this trick.

Why is real data apodized, and what does that cost?

A finite rectangular scan gives a sinc instrument line shape whose sidelobes reach about $\pm22\%$ of the main peak — ripples that can be mistaken for weak spectral lines. Apodization multiplies the interferogram by a smooth taper (triangular, Happ–Genzel, Norton–Beer) that goes gently to zero at the scan ends, which suppresses those sidelobes. The price is a broader main lobe: apodization trades a modest loss of resolution for a much cleaner, artifact-free line shape, and the choice of taper is a deliberate compromise for the measurement at hand.

When does the multiplex advantage disappear?

The Fellgett advantage assumes the noise is additive and independent of the signal — true for thermal infrared detectors, where it delivers its full $\sqrt{M}$ gain. If instead the dominant noise is photon shot noise from a bright source, the noise grows with the total light on the detector; piling all $M$ elements onto one detector then piles up their shot noise too, and the advantage shrinks to nothing or reverses into a multiplex disadvantage. This is why FTS dominates the infrared but competes on more even terms with dispersive and, especially, modern array-detector instruments in the visible and ultraviolet.

Why can an FTS interferogram be asymmetric, and how is that fixed?

Ideally the interferogram is an even function of path difference and a pure cosine transform returns the spectrum. But wavelength-dependent phase errors — from a dispersive beamsplitter, imperfect compensation, or electronic delays — make the autocorrelation complex, so the interferogram is no longer symmetric. Transforming it naively distorts the intensities. The remedy is to record a double-sided interferogram (or a short double-sided plus a long single-sided one) and apply a phase correction, such as the Mertz or Forman method, taking the magnitude of the complex transform so the recovered line intensities are undistorted.