The short answer is: yes, FT8 can actually decode signals that are physically weaker than the overall noise in the receiving channel, but there is an important technical hitch in how we define that noise.
Here is a breakdown of how it works:
In the article you will read
Reference bandwidth (2500 Hz)
When you see in the program WSJT-X data for example -21 dB, this value is not absolute. It is relative to the standard bandwidth of the SSB filter, which is 2500 Hz.
If the signal level is 0 dB, it means that the signal power is the same as the noise power in the entire 2500 Hz band.
If the signal is -21 dB, it means that it is deeply 'drowned' in noise if we looked at it as a whole.
Narrowband gain
However, FT8 does not use the entire 2500 Hz. One FT8 signal occupies only about 50 Hz. Physics does not allow: if you divide the noise from the 2500 Hz band into small 50 Hz segments, in each such small segment there is significantly less noise. By 'focusing' on a very narrow bandwidth, FT8 gains a huge advantage over SSB.
You can listen to how signals sound with different bandwidths at https://olgierd.github.io/ft8-vs-cw/ (TNX for the tip Matej OK1TEH)
Why is FT8 better than CW or SSB?
For intelligibility to the human ear, we need a certain signal-to-noise ratio (SNR). Digital modes like FT8 use techniques that the ear simply cannot:
Forward Error Correction (FEC): Data is transmitted with redundant information. Even if part of the tones is completely lost in noise, mathematical algorithms can calculate and fill in the missing pieces of the message.
Time integration: An FT8 session lasts 15 seconds. The software 'observes' the frequency for a full 15 seconds and looks for known patterns (tones) in it. It's like staring at a very dark photo for a long time until your eyes start to discern outlines.
Bandwidth is just the beginning of the story. What makes FT8 almost 'magical' in pulling signals from the digital grave is a combination of mathematics, precise timing, and very specific modulation..
Here are the key pillars that make FT8 dominate:
Time synchronization and 'Costas Arrays'
Time is critical. FT8 operates in 15-second cycles.
Why is this important? The decoder knows exactly when the signal starts and when it ends. It doesn't have to waste computational power looking for the start of the message in noise.
Costas Arrays: At the beginning, middle, and end of each session, special sequences of tones (synchronization markers) are transmitted. The decoder looks for them like beacons. When it finds them, it knows exactly how to 'align' the rest of the data, even when the signal is extremely weak.
8-GFSK modulation (8 tones)
FT8 uses not just 'a few tones', but exactly 8 frequencies (8-ary Frequency Shift Keying).
Each tone represents 3 bits of data.
G (Gaussian): The tones switch smoothly (through a Gaussian filter), which minimizes side emissions and saves bandwidth.
Constant amplitude: Unlike SSB, where power fluctuates according to voice, FT8 transmits always at 'full throttle'. This allows the final stage of the transmitter to operate in an efficient mode without distortion.
LDPC: Mathematical superhero (Error Correction)
This is probably the biggest reason for success. FT8 uses LDPC (Low-Density Parity-Check) codes.
The message itself (callsign, locator) has only 77 bits.
However, after adding LDPC coding, a total of 174 bits.
What does it do? These extra bits are pure mathematics. If due to noise or interference up to 30 - 40% of tones are lost or damaged, the LDPC algorithm can recover the rest of the message calculate.. It's like a puzzle where you're missing a third of the pieces, but you know exactly what they need to be in order for the picture to make sense.
Structured protocol
Unlike CW (telegraphy), where you can send anything, FT8 is highly structured. The decoder knows in advance that it expects:
Callsign 1
Callsign 2
Locator or report
By having a 'dictionary' of the protocol limited, the probability of successful decoding at low signal-to-noise ratio (SNR) dramatically increases.
Summary: are the signals below the noise level?
From the perspective of measuring signal strength, FT8 mixes pears with apples. The strength is reported at a bandwidth of 2500 Hz, but the decoder works with approx. 50 Hz bandwidth.
A similar effect would be observed when receiving a CW signal - on a wide filter, the signal will be in noise, for a narrower filter it will be better, and for an even narrower filter best.
So don't let anyone pull the wool over your eyes - FT8 signals are not below the noise level.