Part 5 of The Analog Edge, a 14-part field guide to the analog half of a GopherTrunk installation. Part 4 closed with a warning: clearing clipping and intermod narrows the front-end search but doesn’t end it, because the worst analog failure leaves every level statistic pristine. This part is that failure — the one that cost the project its longest hunt, #764, and taught the signature our marginal reader most needs to recognize: carrier-clean but modulation-degraded. If your waterfall looks great and your decoder disagrees, this is the part to read twice.
TL;DR: Your tuner’s local oscillator is multiplied into every sample, so its phase noise — tiny random jitter in when it crosses zero — lands directly on the received phase. On a phase-modulated signal that is the payload. The #764 verdict, from the reporter’s own captures of one site on one antenna: at 2.5 MS/s, demod SNR ≈ 19.7 dB, EVM 7.4%, locks; at 10 MS/s, demod SNR ≈ 9.5 dB, EVM 22.5%, never locks. Neither clips (both peak ≈ −48 dBFS), and the wideband FFT carrier SNR was actually higher at 10 MS/s. An independent-resampler A/B pinned the deficit into the samples themselves — the signature of front-end phase noise / reciprocal mixing at the Airspy’s native 10 MS/s clock. Operator rule: prefer rates your front end is clean at, and never trust an FFT to grade modulation.
Key takeaways
- The LO’s flaws are multiplied into every sample. Mixing doesn’t add the oscillator’s jitter, it transfers it — every received symbol inherits the LO’s phase wander, and no downstream filter can remove what is now part of the signal.
- Carrier-clean but modulation-degraded is the tell. Power metrics, the waterfall, even wideband FFT SNR stay healthy while EVM balloons and demod SNR collapses. It’s the one front-end failure the whole level toolkit of Parts 2–4 cannot see.
- Reciprocal mixing turns neighbors into noise. A noisy LO also smears every strong nearby signal’s energy across your channel — the receiver equivalent of everyone on the band shouting through your oscillator.
- The capture rate can change the analog front end. Same chip, same antenna, different sampling configuration, different oscillator behavior — which is why “works at 2.5 MS/s, fails at 10 MS/s” can be a true statement about hardware, not software.
Cheat sheet
| Concern | What it tells you | Where it lives |
|---|---|---|
| The physics, one page | oscillator jitter → carrier skirts → reciprocal mixing | Phase noise |
| The full detective story | two real bugs, one red herring, one verdict | Ten Megasamples |
| The quality number that matters | demod SNR / EVM from the actual receiver | gophertrunk replay metrics; the rate-invariance test in internal/scanner/ccdecoder/ddc_highrate_test.go |
| Proving it’s the samples | independent-resampler A/B, in depth | Weak-Signal Engineering, Part 12 |
| Rate choice on Airspy hardware | which rates the front end favors | Airspy rate selection |
| Why levels can’t see it | both #764 captures peaked ≈ −48 dBFS | Part 2’s regime table |
In this post
- The symptom that didn’t fit — everything measured healthy, nothing decoded.
- What phase noise actually does — skirts, smear, and reciprocal mixing.
- The verdict, in numbers — the two-capture table.
- Why the wideband FFT lied — integrated power vs symbol accuracy.
- What an operator does about it — rates, A/Bs, and when to stop blaming software.
The symptom that didn’t fit
Rewind the story to where an operator would have stood. Four P25 control channels on one Airspy R2. At 2.5 MS/s, the nearest tap decodes cleanly. To cover all four, the rate goes up to 10 MS/s — and every tap goes dark, including the strong one that just worked. Levels: fine, ≈ −48 dBFS peak, no clipping (Part 4’s tests all pass). Waterfall: the carriers are right there, plainly visible. Two genuine software bugs were found and fixed along the way — a per-tap CPU blowup and a hardcoded channelizer bin count — and the symptom survived them, even in pure offline replay of the captures (#771). Every level instrument in Parts 2–4 said “healthy”; the decoder said no. When your instruments and your outcome disagree that stubbornly, you’re missing an instrument — and the missing one here measures phase.
What phase noise actually does
A receiver tunes by multiplying the incoming spectrum against a local oscillator. In the idealized diagram the LO is a perfect sinusoid — a single infinitely thin spectral line. A real oscillator wobbles: each zero crossing arrives a few picoseconds early or late, at random. In the frequency domain that timing jitter becomes skirts — a pedestal of noise spreading out from the carrier line, described in dBc/Hz at various offsets (the Field Guide entry unpacks the units).
Multiplication transfers those skirts onto everything the receiver hears, and that hurts you twice:
- Directly: your wanted carrier is convolved with the LO’s smear, so each symbol’s phase arrives with the oscillator’s wander added. For a phase-carrying modulation — P25’s C4FM ultimately conveys frequency/phase trajectories; TETRA’s π/4-DQPSK is explicitly differential phase — that wander is indistinguishable from noise in exactly the dimension the data lives in.
- Reciprocally: every strong neighbor in the front end’s view is also convolved with the skirts, and the edges of a strong neighbor’s smear land on your channel as broadband noise. This is reciprocal mixing — a clean signal into a noisy oscillator produces the same in-channel floor as a noisy signal into a clean oscillator. On a busy trunking band with strong carriers everywhere, a modest LO can set your noise floor all by itself.
Note what neither mechanism touches: total power. The energy isn’t removed, it’s relocated — from crisp symbol positions into a blur around them. Every level metric of Parts 2–4 conserves; only symbol-accuracy metrics (EVM, demod SNR, error rate, CRC yield) see the loss.
The verdict, in numbers
Here is #764 reduced to its evidence table — one site, one antenna, the reporter’s own captures, replayed offline through the same code:
| Measurement | 2.5 MS/s capture | 10 MS/s capture |
|---|---|---|
| Peak level | ≈ −48 dBFS | ≈ −48 dBFS |
| Clipping | none | none |
| Wideband FFT carrier SNR | high | higher |
| Demod SNR | ≈ 19.7 dB | ≈ 9.5 dB |
| EVM | 7.4% | 22.5% |
| Control-channel lock | yes | no |
The clincher was the independent-resampler experiment: decimate the 10 MS/s file 4:1 with a resampler that isn’t GopherTrunk’s, feed the result to the proven 2.5 MS/s decode path, and the same ≈ 9.5 dB deficit comes out — so the missing 10 dB is baked into the captured samples, not introduced by our DDC. (One sentence here; the full methodology of proving “it’s the samples” by rate-invariance and independent resamplers is Weak-Signal Engineering, Part 12.) Carrier-clean, modulation-degraded, level-innocent, rate-correlated: that constellation of facts points at the front end’s oscillator behavior at the Airspy’s native 10 MS/s clock — phase noise / reciprocal mixing recorded permanently into every sample.
Why the wideband FFT lied
The most disorienting row of that table is the FFT one: the capture that failed had the higher wideband carrier SNR. It isn’t a paradox once you know what each instrument integrates. An FFT bin sums power — and phase noise conserves power, merely relocating it within and around the channel. A tall, proud carrier in the waterfall is a statement about energy, not about whether that energy’s phase trajectory still encodes symbols. The demodulator, meanwhile, is an interferometer: it compares each symbol’s phase against where a clean symbol would be, and it feels every picosecond of LO jitter. Same signal, two instruments, opposite verdicts — and the decoder’s is the one correlated with reality. This is the series’ Part 2 lesson escalated one level: first we learned dBFS can’t grade quality; now even spectral SNR can’t. Only demodulation-domain numbers grade modulation.
What an operator does about it
You cannot fix an oscillator in YAML, but you have real levers:
- Prefer rates your front end is clean at. The sampling configuration is part of the analog design — clocking, decimation, and PLL settings shift with it. If a lower rate covers your channels, take it; the Airspy rate-selection notes exist because of exactly this history. The decode path itself is rate-invariant (Part 6), so rate choice is purely a front-end decision.
- A/B by capture, not by vibes. Record the same channel at both candidate rates, replay both, compare demod SNR/EVM/lock. Ten minutes, and the answer is durable evidence instead of an impression.
- Weigh oscillator quality when buying. Part of what separates SDR tiers is exactly LO cleanliness under real clocking loads — worth as much as any dB of gain on a phase-modulated trunking band. (Our hardware comparison is the buying-side companion.)
- Keep strong neighbors out of the front end. Reciprocal mixing needs a strong neighbor to reciprocate with; Part 9’s filters reduce the supply.
- Stop blaming software once the signature matches. Carrier-clean, modulation-degraded, reproducible from a capture, correlated with a front-end configuration change: that’s an analog verdict. File the issue with the capture (Part 10) — but point it at the right side of Part 1’s line.
Where this goes next
This part leaned on a claim that deserves its own post: that GopherTrunk’s
decode path treats every capture rate identically, so rate-correlated
symptoms indict the front end. Part 6
opens that up — how both down-converters normalize to one per-protocol
channel rate, when higher rates genuinely help, what they cost, and the two
log lines (decode can't keep up with real time, soapyremote host_drops)
that look like driver bugs and are actually downstream signals.
FAQ
Is phase noise a defect in my SDR? It’s a budget line in every oscillator ever built — the question is degree, and price roughly tracks it. The operational point isn’t “buy perfection,” it’s that phase-noise behavior can differ between configurations of the same device (as #764 showed across sample rates), so measure your rig at the settings you actually run.
How do I check for this signature without lab gear? With the decoder as the instrument. Capture the same channel under both configurations you’re comparing, replay, and read demod SNR/EVM and lock. Carrier visible in the waterfall + healthy dBFS + collapsed demod SNR that follows one configuration is the fingerprint. No spectrum analyzer required — the receiver is a phase-accurate one already.
Could multipath or ISI produce the same “clean carrier, bad decode” look? Yes, and that’s a fair confounder — linear channel distortion also degrades EVM while conserving power. The discriminators: ISI is equalizable (and GopherTrunk’s equalizers recovering most of the loss points that way), tends to follow location and antenna, and doesn’t track front-end configuration. Phase noise follows the hardware configuration and no linear equalizer can undo it. The #764 deficit tracked the sample rate on one antenna — hardware.
Why did this take an independent resampler to prove? Because the obvious A/B — decode the 10 MS/s capture through our own 10 MS/s path — can’t separate “the samples are bad” from “our high-rate path is bad.” Decimating with a third-party resampler and replaying through the already-proven 2.5 MS/s path removes our high-rate code from the loop entirely. The deficit survived, so it lived in the samples. It’s the same self-consistency discipline the project applies to synthetic tests, aimed at hardware.
Does this mean I should always run the lowest rate? No — it means rate is a front-end quality choice, not a free coverage knob. Part 6 gives the full trade: what higher rates buy (more taps per dongle, wideband hunting), what they cost (front-end cleanliness, CPU, USB), and how to verify your hardware’s clean rates empirically.
Series navigation
Part 5 of 14 · ← Part 4: Clipping, Overload & Intermod · Next → Part 6: Sample Rate — The Decode Path Doesn’t Care; the Front End Does