RIGOROUS RE-TEST · STANDARD CHANNEL MODELS · REAL SATELLITE IMAGERY
The contested channel, done properly
The earlier demo used toy impairments and a generous baseline. This version is built to be scrutinized:
real Sentinel-2 satellite imagery (EuroSAT), complex I/Q symbols over Rayleigh & Rician block
fading with receiver channel-state information, a textbook partial-band jammer, and — most importantly —
an honest outage-based digital baseline: a rate-adaptive, capacity-achieving code that drops frames when the
channel can't support its rate. No thumb on the scale.
The honest headline. Against a properly strong classical system, the learned transceiver does not win
everywhere — and that's the credible result. It wins exactly in the hard regime: low SNR,
deep fades, and worst-case jamming. Rate-adaptive digital wins in good channels, where it exploits SNR the
fixed analog code can't. The hard regime is precisely the deep-space / contested-military regime — so this
maps out where the technology belongs, rather than overselling it.
1 · Rayleigh fading — the crossover
learned JSCCrate-adaptive digital (outage-coded)
Below ~6 dB the JSCC wins (at −2 dB, 25.2 vs 24.0 dB,
+1.2 dB); above it, digital pulls ahead by adapting its rate. The JSCC curve is flat and
never outages — it trades peak fidelity in good channels for a guaranteed floor in bad ones. Under milder
Rician K=6 fading the picture is the same with the crossover pushed a little higher (JSCC wins to ~6 dB).
2 · Worst-case partial-band jamming
A jammer with a fixed power budget (JSR 10 dB) at 10 dB SNR chooses the band
fraction ρ that hurts most. Small ρ = a narrow, tall spike; ρ=1 = barrage.
learned JSCCrate-adaptive digital
The jammer's best move against digital is barrage (ρ=1.00), and that is exactly where
the JSCC wins: 22.1 dB vs 21.2 dB. The
JSCC is also far flatter across ρ — the jammer can't find a band fraction that breaks it.
3 · Averages hide lost frames
The digital number is an expected PSNR that averages delivered frames with outages (blank frames).
Under worst-case barrage jamming, digital's 21.2 dB is really ~81% of frames delivered + ~19% frames
totally lost. The JSCC's 22.1 dB is every frame, at consistent quality. For deep-space and
targeting/ISR links — where a lost frame can't be re-sent and a blank frame is useless — that consistency is
worth more than the average PSNR suggests.
4 · See it — satellite imagery, contested channel
Condition: 0dB Rayleigh + 40% partial-band jam (JSR 8dB). Top: original Sentinel-2 tiles. Middle: learned JSCC
(24.8 dB). Bottom: digital (23.7 dB). The JSCC keeps land-cover structure and colour;
the digital reconstruction is blockier and washes tiles out.
Honest verdict & remaining gaps
What holds up: on real satellite imagery, over standard fading with a real jammer and an honest baseline,
the shared-model transceiver is the better choice in the worst conditions (low SNR, deep fades, barrage
jamming) and delivers every frame with no outage — the exact properties deep-space and contested links need.
It also needs no transmitter channel knowledge and no rate negotiation.
What doesn't (yet): it loses in good channels (a fixed analog code can't exploit high SNR — a real
system would switch schemes adaptively). Still a simulation: no Doppler, carrier/timing recovery, hardware
nonlinearity, or a reactive/follower jammer; perfect receiver CSI is assumed; the digital baseline is
information-theoretic, not a specific LDPC decoder. And a neural decoder can still hallucinate detail — for
science/targeting you'd transmit verifiable residuals. This is an honest map of where the idea helps, not a
deployable radio.
