Ok well...I found the following. It's rather technical, but it's a good explanation of how we lose data/resolution in the highlights with compressed RAW. Thom Hogan seems to agree, which had spawned the research. I think it still depends on certain conditions as to wether its relative or not. The complete article is at
http://www.majid.info/mylos/weblog/2004/05/02-1.html
--------
The D70 only offers compressed NEFs as an option, but mercifully they have improved the performance. Ken Rockwell asserts D70 compressed NEFs are lossless, while Thom Hogan claims:
Leaving off Uncompressed NEF is potentially significant--we've been limited in our ability to post process highlight detail, since some of it is destroyed in compression.
To find out which one is correct, I read the C language source code for Dave Coffin's excellent reverse-engineered, open-source RAW converter, dcraw, which supports the D70. The camera has a 12-bit analog to digital converter (ADC) that digitizes the analog signal coming out of the Sony ICX413AQ CCD sensor. In theory a 12-bit sensor should yield up to 212 = 4096 possible values, but the RAW conversion reduces these 4096 values into 683 by applying a quantization curve. These 683 values are then encoded using a variable number of bits (1 to 10) with a tree structure similar to the lossless Huffmann or Lempel-Ziv compression schemes used by programs like ZIP.
The decoding curve is embedded in the NEF file (and could thus be changed by a firmware upgrade without having to change NEF converters), I used a D70 NEF file made available by Uwe Steinmuller of Digital Outback Photo.
The quantization discards information by converting 12 bits' worth of data into into log2(683) = 9.4 bits' worth of resolution. The dynamic range is unchanged. This is a fairly common technique - digital telephony encodes 12 bits' worth of dynamic range in 8 bits using the so-called A-law and mu-law codecs. I modified the program to output the data for the decoding curve (Excel-compatible CSV format), and plotted the curve (PDF) using linear and log-log scales, along with a quadratic regression fit (courtesy of R). The curve resembles a gamma correction curve, linear for values up to 215, then quadratic.
In conclusion, Thom is right - there is some loss of data, mostly in the form of lowered resolution in the highlights.