Window Functions and the Uncertainty Principle

The trade-off is the working constraint of every technology that reads a spectrum. In ultrafast optics it is the design equation of femtosecond and attosecond science: shorter pulses demand broader bandwidth, and the frontier of attosecond XUV bursts from high-harmonic generation — a 2023 Nobel Prize — is a direct assault on the time side of $\Delta t\,\Delta\nu\ge0.441$. In gravitational-wave astronomy the LIGO and Virgo detectors are capped by shot noise and radiation-pressure back-action; they now inject squeezed vacuum — light pushed off the minimum-uncertainty symmetry so phase noise trades down against amplitude noise — to see past the standard quantum limit. Time-frequency analysis as a whole, from spectrograms of speech and birdsong to the wavelets behind JPEG-2000, is the engineering response to the fact that no single window resolves both when and what. In quantum information the inequality underwrites security: data encoded in conjugate bases cannot be measured in both without disturbance, so eavesdropping is detectable. An open thread runs alongside — entropic uncertainty relations bound a sum of Shannon entropies rather than a product of variances, sharper and free of the pathologies of second moments.

Strike a piano key for an instant and the sound is a click — sharp in time but spectrally a wash of every frequency at once. Hold the key forever and you approach a single pure pitch — sharp in frequency but smeared across all time. Real signals live between these poles, and the uncertainty principle states precisely how much spectral spread you must accept to buy time localization. The Fourier transform converts "how fast does $f$ wiggle" into "where is $\hat f$ supported": differentiation amplifies high frequencies, so a sharply peaked $f$ is forced to carry a broad spectrum, as the first figure shows.

The Gaussian earns its central role because it is the fixed shape of the transform, $\mathrm e^{-ax^2}$ mapping to another Gaussian: it divides its "wiggling budget" as evenly as any function can and sits exactly at the bottom of the trade-off curve. Add a chirp — a quadratic phase — and $\Delta x\,\Delta k$ rises above $\tfrac12$, as the second figure shows: the theorem as a landscape with a single minimum at zero chirp.

In quantum mechanics this same figure says a particle cannot have both a definite position and a definite momentum: squeeze the wavefunction into a small box and its momentum spreads wildly; give it a sharp momentum and it spreads over all space. Nothing is needed beyond Fourier conjugacy and the constant $\hbar$ that converts $k$ into $p$ — Heisenberg's principle is not an extra postulate but a corollary of the wave description, true of any Fourier-conjugate pair.

We fix the symmetric Fourier convention throughout; the numerical bound depends on this choice, so it must be stated before any number is quoted. The kernel is $\mathrm e^{-\mathrm i kx}$ with a $1/\sqrt{2\pi}$ prefactor split evenly between forward and inverse transforms.

where: $f(x)$ — the function in the spatial or time domain, assumed to lie in $L^2(\mathbb R)$; $\hat f(k)$ — its transform in the conjugate (wavenumber or frequency) domain; $k$ — the conjugate variable. The symmetric split makes Plancherel read $\|f\|=\|\hat f\|$ with no stray $2\pi$. The "width" of a function is the root-mean-square spread of its probability-like density $|f|^2$ about the centroid — nothing more than the variance of a normalized distribution.

where: $\langle x\rangle$ — mean position, the centroid of $|f|^2$; $\Delta x$ — standard deviation (rms spread) of $|f|^2$. The spread is about the mean, which need not be zero; a shift of $f$ moves $\langle x\rangle$ but leaves $\Delta x$ unchanged. The spectral width is the identical construction on $|\hat f|^2$. By Plancherel the two densities carry equal total weight, so both variances are well defined for any $f\in L^2$.

where: $\Delta k$ — rms spread of the spectrum; $\langle k\rangle$ — mean wavenumber. A modulation $f\mapsto \mathrm e^{\mathrm i k_0 x}f$ shifts $\langle k\rangle$ but leaves $\Delta k$ fixed — the mirror of shift invariance. The central result binds these two spreads by a hard floor: no function, however shaped, can push the product below one half.

where: $\tfrac12$ — the universal lower bound in the symmetric convention. The value is convention-dependent (it becomes $1/4\pi$ for the ordinary-frequency kernel $\mathrm e^{-2\pi\mathrm i\nu x}$); the existence of a positive floor is not. The engine of the proof is the differentiation rule: in the conjugate domain a derivative is the same as multiplication by $\mathrm i k$, which converts a spectral second moment into a statement about $f'$.

where: $f'$ — the derivative of $f$; under $\mathcal F$ it becomes multiplication by $\mathrm i k$. Read the other way, $\widehat{xf}=\mathrm i\,\hat f'$: multiplication and differentiation swap roles, which is exactly why they are conjugate operators. Promoted to operators, $x$ and $-\mathrm i\,\mathrm d/\mathrm dx$ fail to commute — the abstract source of the bound. Multiplying by $\hbar$ turns them into the physical position and momentum operators.

where: $\hat x$ — position operator; $\hat p$ — momentum operator, Fourier-conjugate to $\hat x$; $\hbar$ — reduced Planck constant. The commutator is a pure number times the identity, which makes the bound state-independent. Feeding $p=\hbar k$ into (4) rescales the spectral spread into Heisenberg's relation — the single line where physics enters, through $\hbar$ and nowhere else.

where: $\Delta p=\hbar\,\Delta k$ — the momentum spread. With $\hbar=1.0546\times10^{-34}\ \mathrm{J\,s}$ the bound is negligible in daily life but decisive at atomic scales. The Gaussian saturates every step. A Gaussian of scale $\sigma$ maps to a Gaussian of reciprocal scale $1/\sigma$ — squeeze one, the other spreads by the same factor.

where: $\sigma$ — spatial scale; spectral scale $1/\sigma$. These give $\Delta x=\sigma/\sqrt2$ and $\Delta k=1/(\sigma\sqrt2)$, so $\Delta x\,\Delta k=\tfrac12$ for every $\sigma$ — the equality is scale-invariant. Finally, the discrete face: a window function tapers a finite data record before transforming, trading main-lobe width against side-lobe leakage. The Hann window is the canonical cosine taper.

where: $N$ — number of samples; $n$ — sample index. Tapering the ends to zero removes the discontinuity a rectangular cut leaves, dropping the peak side lobe from $-13$ dB to about $-31.5$ dB at the cost of a main lobe twice as wide — the discrete face of the same trade-off.

The same bound follows from the operator side as a special case of the Robertson relation $\Delta A\,\Delta B\ge\tfrac12|\langle[\hat A,\hat B]\rangle|$, valid for any self-adjoint pair: with $A=\hat x$, $B=-\mathrm i\,\mathrm d/\mathrm dx$ and $[A,B]=\mathrm i\,\mathbb 1$, Cauchy-Schwarz plus keeping the imaginary part gives $\Delta x\,\Delta k\ge\tfrac12|\langle f,\mathrm i f\rangle|=\tfrac12$, and $B=\hat p/\hbar$ returns $\Delta x\,\Delta p\ge\hbar/2$. Because the commutator is a multiple of the identity, the bound is state-independent — the form that generalizes to angular momentum and energy-time. The next derivation asks which states meet it with equality.

The inequality interpolates between two singular endpoints and is saturated by the Gaussian sitting between them.

1 · An electron confined to an atom (fully worked)

An electron is localized to within $\Delta x=0.10\ \mathrm{nm}=1.0\times10^{-10}\ \mathrm m$, roughly an atom's size. Estimate the minimum spread in its momentum and speed and the kinetic-energy scale implied, using $\hbar=1.055\times10^{-34}\ \mathrm{J\,s}$, $m_e=9.11\times10^{-31}\ \mathrm{kg}$.

Step 1. Heisenberg's relation (7) at saturation gives the momentum spread directly: $\Delta p=\dfrac{\hbar}{2\,\Delta x}=\dfrac{1.055\times10^{-34}}{2\,(1.0\times10^{-10})}=\mathbf{5.3\times10^{-25}\ kg\,m/s}$ — the least uncertainty compatible with pinning the electron to one atom.

Step 2. Divide by the mass for a speed spread: $\Delta v=\Delta p/m_e=(5.3\times10^{-25})/(9.11\times10^{-31})=\mathbf{5.8\times10^{5}\ m/s}$, uncertain at half a million metres per second purely because it is confined.

Step 3. The kinetic-energy scale is $E\sim\dfrac{(\Delta p)^2}{2m_e}=\dfrac{(5.3\times10^{-25})^2}{2\,(9.11\times10^{-31})}=1.5\times10^{-19}\ \mathrm J=\mathbf{0.95\ eV}$. That an atom-sized box forces an energy of order an electron-volt — the scale of atomic binding — is why atoms have the size they do: squeezing tighter costs more kinetic energy than the Coulomb attraction can repay. Confining an electron to $0.1\ \mathrm{nm}$ forces $\Delta v\approx5.8\times10^{5}\ \mathrm{m/s}$ and $E\approx1\ \mathrm{eV}$, setting the size of atoms.

2 · The bandwidth of an ultrashort laser pulse (fully worked)

A mode-locked Ti:sapphire laser emits a transform-limited Gaussian pulse of duration $\Delta t=10\ \mathrm{fs}$ (intensity FWHM) centred at $\lambda_0=800\ \mathrm{nm}$. What minimum spectral bandwidth, in frequency and in wavelength, must it occupy?

Step 1. For a Gaussian pulse, the bound in intensity-FWHM widths is the time-bandwidth product $\Delta t\,\Delta\nu\ge\dfrac{2\ln 2}{\pi}=0.441$. A transform-limited pulse saturates it: $\Delta\nu=\dfrac{0.441}{10\times10^{-15}}=4.41\times10^{13}\ \mathrm{Hz}=\mathbf{44\ THz}$.

Step 2. Convert to a wavelength spread with $|\Delta\lambda|=\dfrac{\lambda_0^2}{c}\,\Delta\nu=\dfrac{(800\times10^{-9})^2}{3.00\times10^{8}}\,(4.41\times10^{13})=\mathbf{94\ nm}$ — the pulse must draw on light from roughly $750$ to $850\ \mathrm{nm}$ at once.

Step 3. Read backward: shorter pulses require more bandwidth, which is why femtosecond oscillators use broadband gain media (Ti:sapphire spans $650$ to $1100\ \mathrm{nm}$) and a single-frequency laser can never be pulsed short. The principle is the design equation of ultrafast optics. A $10\ \mathrm{fs}$ pulse at $800\ \mathrm{nm}$ needs at least $44\ \mathrm{THz}$ ($\approx94\ \mathrm{nm}$) of bandwidth.

3 · Resolution versus leakage in a DFT window

Compare the rectangular window with the Hann window on the two competing figures of merit — main-lobe width and peak side-lobe level — for a length-$N$ DFT. The rectangular window is the sharpest cut in time, so it has the narrowest main lobe (null-to-null width $2$ bins, the best raw resolution) but abrupt edges give a Dirichlet transform whose peak side lobe is only $-13\ \mathrm{dB}$ down — one tone leaks conspicuously into its neighbours. The Hann window (9) tapers smoothly to zero, widening the main lobe to $4$ bins (half the resolution) but dropping the peak side lobe to about $-31.5\ \mathrm{dB}$. You cannot have both, because the window and its spectrum obey the same trade-off as $f$ and $\hat f$: separating two equal close tones favours the rectangular window, while detecting a weak tone beside a strong one favours Hann or Blackman.

The uncertainty principle is where nineteenth-century analysis and twentieth-century quantum mechanics turn out to be the same theorem — proved as a fact about Fourier integrals, the physics a corollary of renaming $k$ as $p/\hbar$.

Texts. Stone & Goldbart, Mathematics for Physics, §6 (Fourier methods) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 20 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 13 · Folland, Fourier Analysis and Its Applications, §7.5 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. 1, Complement C-III · Hassani, Mathematical Physics, Ch. on Fourier analysis · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 5. All formulas use the symmetric convention $\hat f=(2\pi)^{-1/2}\int f\,\mathrm e^{-\mathrm i kx}\,\mathrm dx$. Key checkpoints: $\widehat{f'}=\mathrm i k\hat f$ with Plancherel reduces $\int k^2|\hat f|^2$ to $\int|f'|^2$; the Cauchy-Schwarz plus integration-by-parts chain yields $(\Delta x\,\Delta k)^2\ge\tfrac14$; the Gaussian moments $\Delta x=\sigma/\sqrt2$, $\Delta k=1/(\sigma\sqrt2)$ give product $\tfrac12$, and the chirped Gaussian gives $\tfrac12\sqrt{1+(b/a)^2}$. Numerical anchors: the electron ($0.1\ \mathrm{nm}\Rightarrow\Delta p=5.3\times10^{-25}\ \mathrm{kg\,m/s}$, $\Delta v=5.8\times10^{5}\ \mathrm{m/s}$, $E\approx0.95\ \mathrm{eV}$); the pulse ($10\ \mathrm{fs}$ at $800\ \mathrm{nm}\Rightarrow44\ \mathrm{THz}\approx94\ \mathrm{nm}$ via $\Delta t\,\Delta\nu=0.441$); DFT side lobes ($-13\ \mathrm{dB}$ rectangular, $-31.5\ \mathrm{dB}$ Hann).

Prerequisites. The Fourier Transform · Properties of the Fourier Transform · Parseval's and Plancherel's Theorems · The Dirac Delta Function. Leads to. Tempered Distributions and Fourier Transforms · The Convolution Theorem · Solving Differential Equations by Transform Methods.

Ask the tutor: Ask me to run the Cauchy-Schwarz proof line by line, work a Heisenberg estimate for any system you name, or show why your favourite DFT window sits where it does on the time-frequency curve.