ADPCM algorithms work by quantizing the difference between the actual signal and a predicted value. If you decompress the result you will get a signal where the quantization has aldready been performed, and thus you will get the same result if you compress it again.
For example, if you look at an 8-bit sample you will get values that are in the range of 0-255. But if you look at the difference between one sample and the next, most values will be quite small. So you get a great idea: instead of storing each 8-bit sample you store only the first sample as an 8-bit value and after that only the differences as 4-bit signed values. If at any point the difference is greater than what can be stored, you store the maximum and continue from there.
If you now decompress your result,
there will be no two consecutive values where the difference is outside the range -7<=n<=8. This means you can repeat the decompress-compress cycle as many times as you like, and the result will stay the same. However, if you compare the output to your original you will most likely see differences.
(The above is a differential PCM coding algorithm. ADPCM algorithms will predict the next encountered value using some scheme, and store the difference between the predicted value and the actual one. There are some ADPCM algorithms that are lossless, but AFAIK most are lossy.)
You could convert an original high resolution DVD-Audio track into CD-Audio format, but that wouldn't make CD Audio lossy, now would it?
No, but if the source material was anything other than 16-bit 44.1kHz the conversion process would be.