Weakest S/N level detectable ?

Do not be deceived by the statistics of randomly fluctuating quantities.

At signal levels close to the WSPR threshold, successive measurements of S/N will not be identical. This is a measurement issue, and is true even in the absence of actual signal fading. Upward and downward fluctuations are equally likely and will have similar magnitudes when stated as power levels. But when stated in dB, downward fluctuations will be much larger than upward ones.

Suppose the average ratio of signal to noise is 0.0015, with measurement-to-measurement fluctuations of +/-0.0010. (These are perfectly reasonable numbers for a marginal WSPR signal.) Then in dB we have

S/N (upward) = 10*log(0.0025) = -26.0 dB
S/N (average) = 10*log(0.0015) = -28.2 dB
S/N (downward) = 10*log(0.0005) = -33.0 dB

So occasionally you'll see WSPR report S/N values as low as -33 dB, because of measurement fluctuations, even if the "true" S/N was -28 dB.

When the true S/N is as low as -33 dB, the WSPR decoder has virtually no chance of success. You can see this because S/N = -33 dB in 2500 Hz bandwidth means S/N = -33 + 10*log(2500/1.46) = -0.7 dB in the software filters used by the decoder, which have bandwidth 1.46 Hz. Not good!

On the other hand, if S/N = -28 dB in 2500 Hz bandwidth then S/N = -28 + 10*log(2500/1.46) = +4.3 dB, as seen in the decoder's filters. With WSPR's strong forward error correction this is good enough for reliable, error-free decoding.

-- 73, Joe, K1JT