From "wins in the hard regime" toward "never worse, uniquely capable"
Two engineering pushes to strengthen the honest case: a hybrid that recovers good-channel
performance the pure analog code left on the table, and a study of the capabilities digital cannot match when
the channel feedback it relies on isn't available. Reported straight, including how far each got.
Split the 128 complex channel uses into a graceful analog layer (64) plus a digital layer that
codes the residual (64). A digital outage falls back to the analog image, never a blank frame.
~52%of pure-analog's high-SNR gap to digital, recovered
+0.5 dBover digital at 0 dB — and no cliff
pure analog-128pure digital-128hybrid 64+64
A single system that sits on the graceful analog line at low SNR (no cliff, ahead of digital) and
climbs toward digital at high SNR — recovering ~52% of the gap (28.0 vs analog 26.6,
digital 29.3 dB at 18 dB). Honest limit: it does not fully catch digital in a clean channel — the
bandwidth split costs each side — so this is partial dominance, staying within ~1.3 dB of the
best specialist everywhere while never outaging.
Build 2 — the moat: what digital can't do without feedback
Rate-adaptive digital's low-SNR survival depends on a return channel to keep re-choosing its rate.
On a one-way deep-space downlink, a broadcast, or a jammed link where feedback is denied, that luxury is
gone: digital must fix its rate and outages whenever the channel dips. JSCC needs no transmitter channel
knowledge at all.
+8.8 dBJSCC over feedback-free digital at 0 dB
0%of frames delivered by feedback-free digital below its design SNR
JSCC (no TX CSI needed)digital WITH feedback (rate-adaptive)digital WITHOUT feedback (fixed rate)
The dead zone — one-way link at −8 dB
Top: original Sentinel-2 tiles. Middle: JSCC — usable land-cover, no feedback required.
Bottom: feedback-free digital — a blank frame; below its design SNR it delivers nothing.
Honest scope: with a feedback channel, rate-adaptive digital (amber) degrades gracefully at low
SNR too — it isn't dead everywhere. The moat is specifically the feedback-denied regime (one-way, broadcast,
jammed, very-long-delay deep space), where the neural link's zero-CSI operation is a capability digital cannot
cheaply replicate.
Where this leaves the honest case
Never worse, sometimes far better, and uniquely capable. The hybrid keeps the neural link within ~1.3 dB
of the best classical system across the whole SNR range while never cliffing or outaging; and in feedback-denied
links it delivers usable data where digital delivers a blank. Those are defensible claims for exactly the missions
this targets — deep-space downlink, contested/EW, broadcast. The standing limitations are unchanged: simulation-
stage (TRL ~2–3), perfect receiver CSI assumed, information-theoretic digital baseline, and hallucination risk that
argues for verifiable residuals in critical data.
