Every jitter measurement is a comparison between a test clock and a reference clock. In a serial link that reference is reconstructed from the data itself, which decides — entirely — which jitter can cause an error and which cannot.
Jitter is the deviation of a transition from its ideal position. That definition contains a quantity nobody states: ideal according to what. There is no absolute time reference in a communication system, so every jitter measurement is a comparison between the transitions of interest and the transitions of some reference clock, and the answer depends on which reference is chosen.
This is not a philosophical point. Choosing a reference that follows the signal closely makes most of the jitter disappear; choosing one that stands still makes all of it appear. Both numbers are correct measurements of different things, and only one of them predicts errors.
Tracks every movement of the data. The sampling point dances in perfect step with the transitions and no jitter ever causes an error. Impossible to build, and a useful limit to hold in mind.
A clock distributed separately, which ignores what the data does. Every movement of the data is measured in full and every bit of it can cause an error. This is the common-clock case of deck 15, and it is why that scheme ran out of headroom first.
Follows slow movement and ignores fast movement. Jitter below its bandwidth is harmless because the sampling point moves with it; jitter above is not. Everything in this deck follows from where that boundary sits.
In a serial link the reference is extracted from the data stream itself, by locking an oscillator's phase to the observed transitions. It is worth pausing on how odd that is: the timing of the data is being judged against a clock reconstructed from the timing of the data.
That circularity is a feature. Because the recovered clock inherits the data's slow phase movement, that movement cancels and cannot cause an error. It is one of the real advantages of an embedded clock over a distributed one, and it is why a serial link tolerates wander that would destroy a source-synchronous bus.
The loop behaves as a low-pass filter on phase. Below its bandwidth the recovered clock follows the data; above it, the oscillator runs on its own and the data's movement is measured in full. What reaches the sampler is therefore the difference between the data's phase and the clock's, which is the complementary high-pass response.
A transmitter's jitter measured against a laboratory reference and the same transmitter's jitter as seen by a receiver are different quantities, and can differ by an order of magnitude. A specification that does not state the reference — in practice, the loop bandwidth and order of the clock recovery it is to be measured against — has not specified anything testable.
The two curves below are the same loop seen two ways. The jitter transfer is how much of the incoming phase movement the recovered clock reproduces; the error response is what is left over, which is what actually reaches the sampler. They are complements, and their crossover is the loop bandwidth.
Below the loop bandwidth the error response falls at forty decibels per decade, so slow jitter is very strongly rejected: a tone a decade below the bandwidth is attenuated a hundredfold before it reaches the sampler. Above the bandwidth the error response is flat at unity, and every picosecond of jitter is a picosecond of eye closure.
If the answer depends on the reference, the reference has to be standardised. That is what a golden phase-locked loop is: a clock recovery response defined by the standard — a bandwidth, an order, a damping factor — against which compliance measurements are made, regardless of what the actual receiver does.
A physical clock recovery circuit with the specified response, driving the instrument's trigger. Necessary above a few gigabits per second, where a real-time oscilloscope's own bandwidth becomes the limitation, and the only option with an equivalent-time sampling instrument.
A real-time oscilloscope captures the waveform and applies a model of the loop afterwards. More flexible — the same capture can be re-analysed against several standards — and limited by the instrument's own noise and bandwidth rather than by the loop.
The practical consequence is that two laboratories can measure the same transmitter and disagree, not because either is wrong but because they applied different references. When a jitter measurement is disputed, the reference is the first thing to compare and it is usually the explanation.
The random part of the jitter comes from the oscillator, and an oscillator is characterised by its phase noise — the power in a one-hertz bandwidth at a given offset from the carrier, relative to the carrier. Converting that to a jitter in seconds is an integration:
$$\sigma_t = \frac{1}{2\pi f_0}\sqrt{2\int_{f_1}^{f_2} L(f)\,df}.$$
The limits are the point of this slide. They are not a property of the oscillator; they belong to the receiver, because the clock recovery loop decides which offsets matter. Integrating from a lower limit below the loop bandwidth counts phase movement the receiver will track and ignore.
The compliance test that follows from all of this applies a sinusoidal jitter tone at each frequency and asks how much the receiver survives. Only the part the loop fails to track reaches the sampler, so the tolerance is the budget divided by the magnitude of the error response — which is why every published mask has the same shape.
It is periodic jitter at exactly one frequency: bounded, uncorrelated with the data, easy to generate accurately and easy to measure. That makes it useless as a description of anything a real system suffers from and ideal as a calibrated stimulus. A tolerance mask is a statement about the receiver, not about any channel it will meet.
Widening the loop moves the whole tolerance mask to the left, buying tolerance to fast jitter. It is not free: a wider loop also admits more noise from the data transitions into the recovered clock, so the clock itself becomes noisier. Every clock recovery architecture is a position on that trade, which is why standards specify the bandwidth rather than leaving it to the implementer.
Damping is the second lever and it has a visible consequence: an under-damped loop peaks, reproducing more phase movement than it was given over a band just below its bandwidth. In a chain of repeaters or retimers each tracking the last, that peaking accumulates, which is why standards put an explicit limit on jitter transfer peaking as well as on bandwidth.
Clock recovery rejects slow jitter, whatever its cause. It cannot help with jitter that is fast, and the largest single contribution on a lossy channel is both fast and entirely deterministic: the channel's own intersymbol interference moving the zero crossing.
Because this term is a deterministic function of the data, it changes from symbol to symbol — which puts it at the top of the jitter frequency range, where the loop has no effect at all. It is removed by the equaliser, not by the clock recovery and not by a better oscillator, which is the distinction deck 07 argued the decomposition exists to make.
The loop derives its timing from the transitions, and the transitions are the very things being displaced. A channel with heavy data-dependent jitter therefore feeds a corrupted phase reference into the circuit meant to be correcting for it. The equaliser and the clock recovery adapt against each other, which is why they are trained in a defined order and why a link that fails to converge often has nothing wrong with either of them separately.
| Quantity | Expression | Worth remembering |
|---|---|---|
| Rule 3 | — | Every measurement compares a test clock with a reference |
| Jitter transfer | $H(f)$, low-pass | What the recovered clock reproduces; harmless |
| Error response | $1-H(f)$, high-pass | What the sampler sees; falls at 40 dB/decade below the bandwidth |
| Golden PLL | A response defined by the standard | Without it, two laboratories measure different numbers |
| Phase noise to jitter | $\sigma_t=\frac{1}{2\pi f_0}\sqrt{2\int L\,df}$ | The limits belong to the receiver, not the oscillator |
| Tolerance | budget $/\,|1-H(f)|$ | Rises as $f^{-2}$ below the loop bandwidth |
| Peaking | $\zeta < 1$ | Accumulates through a chain of retimers; limited by standards |
| Data-dependent jitter | Fast and deterministic | The loop cannot reject it; the equaliser removes it |
Deck 9 of eleven in Signal Integrity & High-Speed Digital Design. Every figure on this page is computed by si_models/deck09 and embedded as data; nothing is typed in by hand.
Single-page HTML · KaTeX-rendered maths · no build step. Source on GitHub