Time-bandwidth tradeoff

The time-bandwidth tradeoff says that a signal cannot be both narrow in time and narrow in frequency. Make it narrow in one domain, it spreads out in the other.

This is a direct consequence of the Fourier transform’s time-scaling property:

$x(at)\stackrel{\mathcal{F}}{\leftrightarrow }\frac{1}{|a|}X(f/a).$

Compressing in time by factor $a$ (smaller in time) stretches in frequency by the same factor (bigger in frequency). The two width quantities multiplied together stay roughly constant: time-width times bandwidth $\sim$ a constant of order 1.

Canonical examples

The cleanest example: the rectangle and the sinc. A rect of width $w$ in time has Fourier transform a sinc with first zeros at $\pm 1/w$. Narrow rect (small $w$): wide sinc spectrum (first zeros far apart). Wide rect (large $w$): narrow sinc (first zeros close together).

The dual: a sinc of “width” $w$ in time has rect of width $1/w$ in frequency. Same tradeoff.

The impulse $\delta (t)$ — the most narrow possible time-domain signal — has the broadest possible spectrum: $X(f)=1$, the same at every frequency (a “white” spectrum). And dually, a constant in time (infinite width, no localization in time) has spectrum $\delta (f)$ — a single impulse at $f=0$, maximum localization in frequency.

Why it’s fundamental

The same mathematics underlies the Heisenberg uncertainty principle in quantum mechanics: position and momentum are Fourier-conjugate variables, and a particle that’s narrow in position must be broad in momentum (and vice versa). The signals-and-systems version is just the time-frequency form of the same theorem.

A precise version of the tradeoff: for any signal with both a finite spread in time $\Delta t$ and finite spread in frequency $\Delta f$ (measured as standard deviations of $|x(t){|}^2$ and $|X(f){|}^2$), we have $\Delta t\cdot \Delta f\ge 1/(4\pi )$. Equality holds only for Gaussian signals — Gaussians are minimum-uncertainty in this sense.

Practical consequences

Pulse shaping in communications: a short pulse (good time localization, lets you fit many bits per second) requires a broad frequency band. There’s no way around this — narrow pulses need wide spectra. The standard tradeoff: a bit duration $T_b$ requires a bandwidth of order $1/T_b$.

Filter design: a sharp frequency cutoff (narrow rolloff region in frequency) requires a long impulse response in time. A causal filter with brick-wall stopband would need an infinitely long impulse response (which is non-causal, but the principle is the same in the finite case). Sharper cutoff ⟹ longer time response.

Anti-aliasing filters: in Sampling applications, the anti-alias filter has to be sharp enough to block aliases without distorting the signal much. There’s a tradeoff between transition-band width (cleanliness of the cutoff) and filter length (delay introduced).

Smoothness and decay rate

A related principle: smoothness in one domain corresponds to fast decay in the other. A signal that is infinitely differentiable in time (very smooth) has a spectrum that decays faster than any power of $f$. A signal with jumps has a spectrum that decays only as $1/f$. A signal with kinks (continuous but discontinuous derivative) has spectrum decaying as $1/f^2$.

The lesson: rough features in time = slow decay in frequency = lots of high-frequency content. Smooth signals are bandlimited in practice; rough signals are not.