The formula
There are two sinc functions in the wild, and mixing them up costs you a factor of π. Know which one your tool means before you trust a number.
Unnormalized (math convention): sinc(x) = sin(x) / x
Normalized (signal-processing): sinc(x) = sin(π·x) / (π·x)
- Reading it: it's a sine wave divided by its own argument, so the oscillation stays but the amplitude decays like 1/x. The waveform rings and fades.
- At x = 0 both forms are defined by the limit,
sinc(0) = 1. The 0/0 is removable — nothing blows up, and any correct implementation returns exactly 1 there. - Zero crossings are where the numerator's sine hits zero but the denominator doesn't. For the normalized form that's every nonzero integer (±1, ±2, ±3, …). For the unnormalized form it's every nonzero multiple of π. That integer-zero property is the whole reason the normalized version exists.
The transform pair is the part engineers actually use:
FourierTransform{ sinc(t) } = rect(f) (normalized form)
- Conventions, stated: here
rect(f) = 1for |f| < 1/2 and 0 outside (unit width, unit height), and the transform is the Hz convention,X(f) = ∫ x(t)·e^(−j2πf t) dt. Use the ω (rad/s) convention instead and the cutoff moves to ±π — the same factor-of-π trap as the two sinc definitions. - Scaled for a real bandwidth B:
2B·sinc(2Bt) ↔ rect(f/2B), a brick-wall low-pass from −B to +B. That's the form you actually plug numbers into. - Reading it: a sinc in time is a perfect rectangle in frequency, and vice versa. A brick-wall filter in the frequency domain is a sinc in the time domain. That single fact drives everything below.
Two integrals worth memorizing (the first is the Dirichlet integral, Si(∞) = π/2 doubled) [1]:
∫ sin(x)/x dx = π (over all x, the Dirichlet integral)
∫ sinc²(x) dx = 1 (normalized form, over all x)
Where you meet it
The reconstruction filter behind every DAC. When you convert samples back to a continuous waveform, the ideal reconstruction kernel is a sinc: the Whittaker–Shannon formula rebuilds a band-limited signal as a sum of shifted, scaled sincs, one per sample. On the bench you never see the ideal — you see its zero-order-hold approximation and a droop-correction filter cleaning up after it. The sinc is the thing your real hardware is trying and failing to be.
The far-field pattern of a uniform aperture. Illuminate a slot, a horn mouth, or a uniformly excited linear array with constant amplitude, and the radiation pattern is a sinc in the sine of the angle. This is why a uniformly fed aperture has a first sidelobe fixed at roughly −13.3 dB no matter how many elements you add — that number is baked into sin(x)/x, not into your array size. On the antenna range you read that −13 dB off the pattern cut and know instantly you're looking at uniform illumination.
Spectral leakage on the analyzer. Take an FFT of a tone that doesn't land exactly on a bin, and the energy smears into a sinc-shaped skirt. The rectangular window you implicitly applied by grabbing a finite record has a sinc spectrum, and it convolves with your tone. Every window function — Hann, Hamming, Blackman — is a deliberate trade against that sinc's sidelobes.
Bandlimiting and anti-aliasing in review. When someone at a design review claims a filter is "ideal low-pass," the honest follow-up is: then its impulse response is a sinc, which is infinitely long and non-causal, so it does not exist. That's the moment to talk about what you're actually building instead.
How it works
The sinc is the Fourier twin of the rectangle, and that duality is the source of both its usefulness and every trap it sets.
It decays, but slowly. The envelope falls off like 1/x. That's fast enough to converge but slow enough to hurt. The tails matter. A sinc-shaped filter response that looks "basically zero" out past a few lobes is still leaking, and in a spectrum plot those far sidelobes are exactly the dynamic range you were trying to protect.
The sidelobe level is a constant. For the sin(x)/x shape, the first sidelobe peak sits at x ≈ 4.4934 (the first nonzero solution of tan(x) = x), where the value is −0.2172. In power that's 20·log10(0.2172) ≈ −13.26 dB. This is why uniform illumination always gives a −13 dB first sidelobe and uniform-record FFTs always leak at −13 dB relative to the peak. You don't fix it by making the aperture bigger or the FFT longer — bigger only narrows the main lobe. You fix it by tapering the illumination or windowing the record, which trades a wider main lobe for lower sidelobes.
It is not absolutely integrable. The integral of |sin(πx)/(πx)| over all x diverges. The ∫ sin(x)/x = π result only holds as an improper (conditionally convergent) integral — the positive and negative lobes have to cancel in order. Practically: you can't truncate a sinc filter and expect graceful behavior. Chop the tails and you get Gibbs ringing at the band edge, the same overshoot you see on a square wave reconstructed from too few harmonics. Real designs multiply the truncated sinc by a window to tame it, which is windowed-sinc FIR filter design in one sentence.
The mistake people make is the π. NumPy's numpy.sinc and MATLAB's sinc both use the normalized form, sin(πx)/(πx), with zeros at the integers. A textbook or a DSP paper may hand you the unnormalized sin(x)/x. If you feed one convention into code expecting the other, your zero crossings land in the wrong place, your filter cutoff shifts by π, and your reconstruction is quietly wrong. Check the docs before you trust the function.
The other trap is causality. The ideal sinc kernel is symmetric about t = 0 and infinite in both directions, which means it responds before the input arrives. No physical filter can do that. Every real "sinc" filter is a delayed, truncated, windowed approximation, and the delay is the price of getting close.
History
The shape is older than its name. Émile Borel touched the underlying interpolation idea in 1898, but the cleaner story starts with Edmund Taylor Whittaker, who in 1915 published "On the Functions Which are Represented by the Expansions of the Interpolation-Theory" and named the reconstruction the cardinal series — the smoothest band-limited function passing through a given set of samples [2][3]. His son, J. M. Whittaker, carried the theory forward through the 1930s [2].
The idea landed in engineering through Claude Shannon, who in his 1949 paper on communication in the presence of noise stated the sampling theorem in the form engineers still use and cited Whittaker's cardinal series as the reconstruction [2][3]. That's why the interpolation formula carries both names — Whittaker–Shannon — and why the sampling theorem is sometimes written Whittaker–Nyquist–Shannon.
The word "sinc" itself came from British radar work. Philip M. Woodward, with I. L. Davies, introduced the notation in a 1952 paper on information theory applied to telecommunication and radar, remarking that the function turned up so often in Fourier analysis that it deserved a symbol of its own [4][5]. He fixed it in print the next year in his monograph Probability and Information Theory, with Applications to Radar [4][5]. "Sinc" is a contraction of the Latin sinus cardinalis — cardinal sine — tying the name straight back to Whittaker's cardinal series. So the function came out of interpolation theory, got its engineering job from sampling and radar, and got its four-letter name from a man trying to save ink on a page of Fourier integrals.
Related tools
- /tools/antenna-gain-beamwidth
- /tools/fresnel-zone
- /tools/rc-filter
- /tools/adc-resolution
- /tools/db-converter