Symbol Timing, Carrier Recovery & Inter-Symbol Interference

A digital receiver sees samples from its own clock and oscillator, not the transmitter's. It must infer where symbols peak, how fast their phase rotates, which constellation orientation is correct, where frames begin, and how earlier symbols leak into the current one. These are coupled estimation problems: a badly timed sample looks like channel distortion, ISI biases timing and carrier detectors, and a carrier loop can lock to the wrong rotational ambiguity before framing exposes it.

The receiver should therefore separate acquisition from tracking. Acquisition searches a wide uncertainty region using a known preamble or cyclostationary feature. Tracking follows small residual drift with a narrow loop. One gain setting rarely performs both jobs well.

What the receiver does not know #

A useful complex-baseband model is

$$r(t)=A e^{j(2\pi\Delta f t+\phi)} \sum_k a_k g(t-kT-\tau)+w(t),$$

where $A$ is gain, $\Delta f$ carrier-frequency offset, $\phi$ carrier phase, $T$ the symbol interval, $\tau$ timing phase, $g(t)$ the combined transmit-pulse/channel/receive-filter response, and $w(t)$ noise. Frame position and sampling-frequency offset are additional unknowns: the latter makes $\tau$ drift instead of remaining constant.

The ordering is not universally fixed, but a robust burst receiver usually removes gross frequency error before a long matched-filter integration, obtains coarse timing and gain from a preamble, then enables narrow timing/carrier tracking and finally frame/channel decoding. Each handoff needs an explicit lock criterion.

ISI is sampled channel memory #

After the receive filter, sampling at $t=mT+\tau$ gives

$$y[m]=\sum_k a_k h[(m-k)T+\tau]+n[m].$$

The desired symbol contributes $a_m h(\tau)$; all $k\ne m$ terms are inter-symbol interference (ISI). At the correct phase, a zero-ISI pulse satisfies the first Nyquist criterion

$$h(nT)=\begin{cases}1,&n=0\\0,&n\ne0.\end{cases}$$

The sinc pulse satisfies this exactly but is infinite in time and unrealizable. A raised-cosine (RC) response preserves the zero crossings while introducing roll-off $0\le\alpha\le1$. Its one-sided baseband bandwidth is

$$B=\frac{1+\alpha}{2T}.$$

Practical links split the response into root-raised-cosine (RRC) filters at transmitter and receiver; their product is RC. Two matched RRC filters therefore create the intended zero-ISI samples only when their rate, roll-off, truncation, normalization, and sample phase agree and the physical channel adds no uncompensated distortion.

Timing recovery is a feedback estimator #

A timing loop repeatedly performs four operations:

1. an interpolator produces samples at a fractional time chosen by the current estimate;
2. a timing-error detector (TED) converts local waveform shape into an error $e_k$;
3. a loop filter trades acquisition speed against jitter/noise rejection;
4. a numerically controlled oscillator advances the next sample phase and symbol period.

flowchart LR
  X[matched-filter samples] --> I[fractional interpolator]
  I --> TED[timing-error detector]
  TED --> LF[loop filter]
  LF --> NCO[phase and rate accumulator]
  NCO --> I
  I --> Y[symbol-rate samples]

The mean TED output versus timing offset is its S-curve. The desired lock point is a zero crossing with the correct restoring slope. That slope is the TED gain; if amplitude normalization, pulse shape, samples/symbol, or $E_s/N_0$ differs from the assumed value, the implemented loop bandwidth and damping differ too. GNU Radio's Symbol Sync documentation explicitly treats expected TED gain as a sensitive parameter for this reason.

Common TEDs make different assumptions:

Worked ISI cancellation #

Consider a symbol-rate channel

$$y[k]=a[k]+0.45a[k-1]+n[k],\qquad a[k]\in\{-1,+1\}.$$

For transmitted symbols $a=[+1,-1,-1,+1]$ and no noise, taking $a[-1]=0$ gives

A zero-threshold slicer still succeeds here, but the margin on symbols 1 and 3 has fallen from 1 to 0.55. A decision-feedback equalizer (DFE) forms

$$z[k]=y[k]-0.45\hat a[k-1].$$

If the previous decision is correct, $z[k]=a[k]+n[k]$ and the postcursor is cancelled without enhancing noise. If one decision is wrong by 2 in amplitude, the next sample receives a feedback error of $0.45\times2=0.9$—large enough to cause error propagation. This is the DFE trade: no linear noise enhancement, but hard dependence on previous decisions.

Carrier acquisition and tracking #

With residual frequency offset $\Delta f$, the constellation rotates by

$$\Delta\theta=2\pi\Delta fT$$

per symbol. Over a matched-filter interval, a large offset also causes coherent-integration loss; it is therefore not always safe to put coarse frequency correction after the matched filter.

A carrier loop uses a phase detector, loop filter, and NCO. For BPSK/QPSK a Costas-loop family removes data modulation from the phase error; decision-directed detectors extend the idea to QAM after decisions are reliable. Pilots or a known preamble can estimate phase without decision feedback.

Symmetric constellations create ambiguities: QPSK can lock at rotations separated by $\pi/2$. Differential coding, pilot polarity, or frame fields resolve the ambiguity. A tight-looking constellation is not proof of correct absolute rotation.

Loop bandwidth makes the same bargain as in timing recovery. Wider bandwidth acquires and follows drift faster but admits more noise and reference modulation. Narrower bandwidth reduces jitter after lock but can lose a burst, Doppler ramp, or oscillator transient. Many receivers explicitly switch gains after a lock detector asserts.

Frame synchronization is a multiple-hypothesis test #

For a known length-$L$ marker $s_i\in\{-1,+1\}$, a normalized sliding correlation is

$$\rho[m]=\frac{\left|\sum_{i=0}^{L-1}r[m+i]s_i\right|^2} {\left(\sum |r[m+i]|^2\right)L}.$$

Normalization makes the threshold less sensitive to gain. It does not eliminate the false-alarm trade: searching $M$ candidate positions turns a per-cell false-alarm probability $p$ into

$$P_{FA,total}=1-(1-p)^M\approx Mp \quad (p\ll1).$$

Long searches therefore need stronger thresholds, confirmation across repeated markers, or structural checks after the candidate boundary. CCSDS Attached Sync Markers are one concrete example of treating the marker, randomizer state, codeword boundary, and decoder synchronization as one ordered contract.

Equalizers choose where to pay #

An equalizer cannot repair a receiver whose sample clock is slipping beyond its adaptation model. Conversely, a timing loop can be biased by an unequilibrated asymmetric channel. Training commonly alternates or coordinates their adaptation rather than assuming independent convergence.

A reproducible acquisition test #

Do not validate the chain only by final BER. Record, against the same absolute sample index:

• coarse frequency estimate and residual phase slope;
• interpolator phase, estimated samples/symbol, TED error, and loop lock flag;
• equalizer taps and error power;
• carrier phase/frequency state and detector error;
• frame-correlation peak, runner-up peak, threshold, and chosen boundary;
• pre-decoder symbols or LLRs and post-decoder frame/CRC result.

Then exercise deterministic offsets at both capture and tracking limits: positive and negative carrier offset, sampling-rate error, fractional-symbol timing, gain/DC error, multipath delay, burst start, and phase ambiguity. A loop is verified only when it acquires from the declared uncertainty region, remains locked through the declared dynamics, and fails visibly outside them.

The correct mental model is one coupled state machine: search → coarse frequency/gain → timing acquisition → channel estimate/equalization → carrier tracking → frame confirmation → payload tracking. The exact order may change, but every transition needs evidence and every estimator needs an explicit operating range.