Loading 3D assets...
...
Drag to orbit - wheel / pinch to zoom - the knobs turn with the sliders; the screen overlays many two-symbol waveforms into an eye

Eye Diagram Lab: Raised-Cosine Pulses, Intersymbol Interference and Bit Errors

Send random bits as pulses and overlay the received waveform on an oscilloscope: the 'eye' shows how much margin a receiver has. A raised-cosine pulse has zero intersymbol interference at the right sampling instant; sample early or late, add noise or timing jitter, and the eye closes and bit errors appear. Pure signal-processing model: no data is translated or decoded as language.
Main-cursor sample p(offset)-
Worst-case ISI: sum of |p(offset - k)|-
Eye height, noise-free (1 = ideal)-
Eye opening-
Bit error rate (analytic)-
Bit error rate (Monte Carlo)-
Occupied bandwidth (1+β)/2-
Best sampling offset-
Eye diagram: about 110 overlaid two-symbol traces. Amber dashed line: sampling instant; red ticks: noise-free worst-case +/- levels there.
log10 of the bit error rate versus sampling offset: blue without jitter, red with your jitter, yellow dot the Monte Carlo result at your offset.
The pulse p(t) in symbol periods. Green dot: the wanted sample; yellow dots: the neighbouring pulses' contributions (ISI) at your sampling offset. At zero offset they all sit on zero crossings.
Independent numerical checks
Each line compares two separate calculations; a FAIL would show here.
The science

What is modelled. Binary symbols ak = ±1 are sent once every T as pulses, so the received waveform is s(t) = Σk ak p(t - kT). The pulse is the raised cosine p(t) = sinc(t/T) cos(πβt/T) / (1 - (2βt/T)2) (a limit value is used where the denominator vanishes). Its defining property is the Nyquist criterion: p(kT) = 0 for every integer k ≠ 0, so if the receiver samples exactly at the symbol instants the neighbours contribute nothing. Roll-off β trades bandwidth, (1+β)/(2T), for a pulse that decays faster (as 1/t3 for β > 0, but only 1/t for the sinc at β = 0), so a timing error costs far less at larger β.

Eye and errors. Sampling at offset τ gives y = a0 p(τ) + Σk≠0 ak p(τ - k) + n with Gaussian noise n of standard deviation σ. The peak-distortion eye height is p(τ) - Σk≠0 |p(τ - k)| (the worst bit pattern). The bit error probability is the average over all neighbour patterns of Q((p(τ) + ISI)/σ), where Q is the Gaussian tail; timing jitter is averaged with a 7-point Gauss-Hermite rule. Neighbours up to ±5 symbols are included (the tails beyond that are tiny for β > 0 and still small for the sinc).

Checks. (1) All integer-time zero crossings (Nyquist). (2) The pulse and its shifts sum to 1 (sampling theorem). (3) The closed-form eye height against a brute-force search over all 1024 bit patterns. (4) The erfc routine at Q(3) against a Simpson integral of the Gaussian density. (5) The analytic bit error rate against a Monte Carlo of random bits with Gaussian noise and jitter, within statistical tolerance.

Not modelled. A real channel filter and matched receiver filter (the pulse is taken as the full end-to-end response), non-Gaussian and coloured noise, clock recovery, equalisation, multi-level modulation, finite-length truncation of the pulse, and nonlinearities. The noise drawn on the eye is per-sample for appearance only; only the noise at the sampling instant enters the error rate. This page says nothing about speech, language or translation: it only models how a binary signal is sampled. Formulas are standard textbook results quoted from memory.

The oscilloscope, pulse generator, knobs, connectors and bench mat were modelled and Cycles-baked in Blender; the pulse maths, the brute-force and Monte Carlo checks and the live screen texture run in Three.js r160.

Made with 🟠 Blender · Three.js