No reason for confusion.

The recording systems makes no judgements on what is wanted signal and what is noise. It attempts to record everything as well as it can. How could it tell what is dialog and which is unwanted machinery noise, for example?

Noise in this context is not the noises we hear on set which we try to avoid recording (that is another matter more to do with miking practises, choosing the location etc), but noise caused by the less than perfect recording system: mics, mixer, mic preamplifiers and AD converter. It is all electronically generated and we hear it only on the recording or when monitoring it.

Sampling bit depth, either 16 or 24 bits, dictates the accuracy of the samples. Simply put, how small are the smallest voltage variations in the incoming signal which can be recorded. AD converter system measures voltages, for that reason one additional bit can describe twice as big voltage differentials and thus one bit more gives 6.03 dB more dynamic range potential to the system.

There is no other measurement tool for this accuracy, only the difference between smallest and largest possible voltage (waveform) and the relative logarithmic unit dB is used. If you want to use "normal" number ratios, you just raise 10 to dB/10 power (10^(dB/10)).

Dymnamic range does not mean only the difference between the quietest sound and loudest in a given recording, but also the difference between the smallest detail and the general level of the signal. Remember that the sound wave is very complex single waveform which contains all the levels and frequences in one single wave.

Now, let's imagine a perfect clean mic signal with 144 dB dynamic range going to a cheap 24 bit recorder. Cheap electronics add so much noise, that the finest 48 dB/8 bits of detail is buriend under it. We have 96 dB of clean signal and if we listen loud enough (this is important, we never do...), we hear just noise under this good signal. Feeding this same signal to a 16 bit recorder we get the same 16 bits worth of good signal, and if we turn up the volume, we hear just noise just like before. Both recordings are of same quality.

The misunderstanding about this "resolution" and dynamic range thing is that people forget that this dynamic range is the relative measurement of the smallest detail level piggybacking the large amplitude waveforms. If the dynamic range is large (like in good 24 bit systems), then there is more "resolution". If the noise level is high, then in a 24 bit signal these minuscule details piggybacking the waveform is not real signal, but random noise. It can not be called "detail" or "resolution".

Like I said before, if just using more bits for sampling would really improve the signal quality regardless the noise floor provided by the less than perfect analog circuits, we would certanly be using 48 bit systems by now. But we are not. That is because there is no point in using more bits than what the other parts of the chain can feed. With the best analog electronics and microphones the limit is around 130 dB which translates to 22 bit sample depth (For this reason 24 bit AD converters are plenty good enough and nobody even tries to use more).

Back to cheap recorders: with high noise electronics these recorder's 24 bit AD converters are not fed with a signal which has more than 96 dB or so dynamic range. So there is NOTHING those 8 extra bits compared to 16 bit recorders can do exept record this random noise faithfully. For that reason 24 bit files have exactly the same data content in the top 16 bits as a 16 bit file of the same signa, and ONLY noise in the last 8 least significant bits. That makes using 24 bit recording pointless, UNLESS the electronics & mics actually are capable of more than 96 dB dynamic range. With most cheap recorders they are not.

And finally, to put this all into perspective: 16 bit recording is wastly better than any analog recorders are capable of. Tens of thousands of high quality movies have been recorded in not-so-distant past using $15000 analog Nagra recorders, with ONLY about 65 dB dynamic range (11 bits...). Now, amateurs are panicing if they are not getting true 24 bit resolutions = 144 dB dynamic range from their $300 pocket recorders! Come on, in XH-A1 we have qualitywise a much better audio recorder than those glorified Nagras. That surely can not be the reason why we can not make great movies with great audio?

1.
If we have two BW pictures - one in 8 bit resolution and the some with 24 bit resolution. Both have he some dark poinst at zero black level, brithness and full white points. Both pictures have the some dynamic range. But the 24bit picture has 24 million levels between black and white - 24millions of grey levels. 8 bit picture has only 256 levels. This is the some as a difference between recorded audio signal.

2.
To be exactly - dynamic range is difference between max and min possible writable level.
By 24 bit resolution - we can write 0000.....001 and this - lovest writable level will be 1/24million of possible max level. By 8 bit resolution we have 00000001 and this is 1/256 of possible max level. The lovest level about zero is 366 times bigger than by 24 bit! This is the difference. Lower first step = greater dynamic range. Finer details just about zero = greater dynamic range. If we have not a really deep black in our pictures - because dark areas are noisy - we have pictures with lover dynamic range - but the 24bit picture has still finer grey graduations.
The some in audio. We record to max level - to 0dB. Every 6dB to minus is a half of the level. Logarithmic. -24dB ist a 1/16 of max. possible level. -30dB 1/32 of maximal level.
What if we have a ground noise level on -30dB? We have very low dynamic range - only 30dB. But how much possible levels - how fine steps are about noise level? We can`t use 1/32 of all levels - because this area is only noise. 24millions/32= 750 000. This is what we can`t use. Usable area has still 23,25 million levels.
Horrible dynamic range - but still much greater quality due to excellent resolution above the noise level.
Its mistake to say - we record with only 30dB dynamic range - so we have only ~8bit resolution and 16 bits are only noise. The difference in dynamic range makes the "quality" in lovest audio levels - but better resolution works in whole area - even to loudest 0dB.
Audio resolution and audio levels are linear - but dynamic range logarythmic and we have to different this.
3. Back to the point here.
We record audio - but not to 0dB to avoid oversteering by unexpected louder sounds.
We record to -6dB - to half of a full possible level. We record a real life sounds - probably nothing will have a dynamic about 60-70 dB. Yes - for this dynamic range in full levels will be 16 bit still more than enough. But by recording to -6dB we use only a half of 16 bit resolution!
In postproduction we have to do sound levels twice as louder - to have in final levels to 0dB. So - we make all sound louder - even noise = it will be no better dynamic range. But even quantisation steps will be twice as bigger! This is the some as scaling to different range here:Resolution Compromise 8 to 12-bit
We have here a final sound - with twice as big quantisation steps as by recordings directly to 0dB! And this is the problem.
If we record in 24 bit to -6dB an makes normalising to 0dB - every quantisation step is still 366times finer. After downconversion to 16bit - the is much better quality.
Even postproduction makes sense to use 24bit recording. This is still not enough. Probably ever better music software uses 32bit resolution and temporary files for every processing -even on 16bit files.

Absolutelly NO. Dynamic range is not the only one aspect of quality.
There is definitelly no problem today for any recorder to have true 24 bit resolution. The problem is on analogue devices side on recording equipments. Its no a technical problem to produce smalest 24bit ADC with an precise 1bit technology - but its a problem to deliver them analogue signals with 144dB dynamic range. Its a problem to find any signals wit 144dB dynamic range.
If minimal noticeable for peoples ear sounds are by 0dB - a jet airplane products still only 120dB about this! A normally conversation is at 60dB. Remember - this is a logarithmic range.
120dB is not twice as loud as 60dB. 66dB is it. 72dB is 4 times louder as 60dB.
So 144dB is 16 times louder as a jet airplane with 120dB!
So - why in the hell would anybody produce a recorder with 24 bit if there were nothing more as a dynamic range? To record simultaneously lovest sounds and 16 starting Jets?
If we had ability (mics preamps and recorders) to record simply even half of this dynamic range - the audio level regulators are useless.

This analogy is invalid. In audio having more sample depth means the quetest parts can be more quiet. In photography this would mean that blacks would be blacker. Zero blackness is called "black body" in physics, theoretical term impossible to achieve in reality. If your photograph would be analogous with audio, black in the the 24 bit photo would have a reflectance of 0.00000596% and in 8 bit photo 0.39%. Quite a difference. In photography first the dynamic range is set governed by the sensor quality (typically 7 to 12 bit equvalency = 42 to 72 dB dynamic range), then this range is quantified with certain accuracy, typically 8 bits, sometimes 14 or 16. In audio one bit is always 6 dB.

Case 2 is a complete mess. Hopeless. You have the math upside down. Fine gradiations in the signal are the bottom part of the dynamic range. If it is noise, it is noise.

There is no other measure or unit of this quality than dynamic range. It all boils down to dynamic range: ratio between the total maximum amplitude and the finest details in that waveform.

24 bit recorders are made for 2 reasons: 1) because we can, and: 2) using 24 bit sample depth with good front end gives more latitude and safety in recording when aiming for maximum quality end result, even if the said result is usually 16 bits (which in real life is much better than the reproduction systems and listening rooms).

Thanks Petri.
Generally we have finest details in 24 bit vaweform and lover quantisation noise. So it still make sense to use 18, 20 or 24 bit if we record to -6dB and normalize in postproduction. Because recording to -6dB we use only a half of the maximum amplitude and we add less quantisation noise to the noise of analog side.